Add files using upload-large-folder tool

fla/models/abc/configuration_abc.py

@@ -0,0 +1,91 @@
+# -*- coding: utf-8 -*-
+
+from typing import Dict, Optional
+
+from transformers.configuration_utils import PretrainedConfig
+
+
+class ABCConfig(PretrainedConfig):
+
+    model_type = 'abc'
+    keys_to_ignore_at_inference = ['past_key_values']
+
+    def __init__(
+        self,
+        hidden_size: int = 2048,
+        gate_low_rank_dim: int = 16,
+        clamp_min: float = -32,
+        clamp_max: float = 32,
+        hidden_ratio: Optional[int] = 4,
+        intermediate_size: Optional[int] = None,
+        num_hidden_layers: int = 24,
+        num_heads: int = 4,
+        num_slots: Optional[int] = 64,
+        use_short_conv: bool = False,
+        conv_size: int = 4,
+        exapnd_k: float = 0.5,
+        exapnd_v: float = 1,
+        hidden_act: str = "swish",
+        max_position_embeddings: int = 2048,
+        elementwise_affine: Optional[bool] = True,
+        norm_eps: float = 1e-6,
+        use_rope: bool = True,
+        attn: Optional[Dict] = None,
+        use_cache: bool = True,
+        pad_token_id: int = None,
+        bos_token_id: int = 1,
+        eos_token_id: int = 2,
+        tie_word_embeddings: bool = False,
+        initializer_range: float = 0.006,
+        fuse_norm: bool = True,
+        fuse_swiglu: bool = True,
+        fuse_cross_entropy: bool = True,
+        vocab_size: int = 32000,
+        **kwargs
+    ):
+        self.hidden_size = hidden_size
+        self.gate_low_rank_dim = gate_low_rank_dim
+        self.clamp_min = clamp_min
+        self.clamp_max = clamp_max
+        self.hidden_ratio = hidden_ratio
+        self.intermediate_size = intermediate_size
+        self.num_hidden_layers = num_hidden_layers
+        self.num_heads = num_heads
+        self.num_slots = num_slots
+        self.use_short_conv = use_short_conv
+        self.conv_size = conv_size
+        self.expand_k = exapnd_k
+        self.expand_v = exapnd_v
+        self.hidden_act = hidden_act
+        self.max_position_embeddings = max_position_embeddings
+        self.elementwise_affine = elementwise_affine
+        self.norm_eps = norm_eps
+        self.use_rope = use_rope
+        self.attn = attn
+        self.use_cache = use_cache
+        self.initializer_range = initializer_range
+
+        self.fuse_norm = fuse_norm
+        self.fuse_swiglu = fuse_swiglu
+        self.fuse_cross_entropy = fuse_cross_entropy
+        self.vocab_size = vocab_size
+
+        if attn is not None:
+            if not isinstance(attn, Dict):
+                raise ValueError("attn must be a dictionary")
+            if 'layers' not in attn:
+                raise ValueError("Layer indices must be provided to initialize hybrid attention layers")
+            if 'num_heads' not in attn:
+                raise ValueError("Number of heads must be provided to initialize hybrid attention layers")
+            attn['num_kv_heads'] = attn.get('num_kv_heads', attn['num_heads'])
+            attn['qkv_bias'] = attn.get('qkv_bias', False)
+            attn['window_size'] = attn.get('window_size', None)
+            attn['rope_theta'] = attn.get('rope_theta', 10000.)
+
+        super().__init__(
+            pad_token_id=pad_token_id,
+            bos_token_id=bos_token_id,
+            eos_token_id=eos_token_id,
+            tie_word_embeddings=tie_word_embeddings,
+            **kwargs,
+        )

fla/models/abc/modeling_abc.py

@@ -0,0 +1,418 @@
+# -*- coding: utf-8 -*-
+
+from __future__ import annotations
+
+import math
+import warnings
+from typing import TYPE_CHECKING, Dict, List, Optional, Tuple, Union
+
+import torch
+import torch.nn as nn
+import torch.utils.checkpoint
+from transformers.generation import GenerationMixin
+from transformers.modeling_outputs import BaseModelOutputWithPast, CausalLMOutputWithPast
+from transformers.modeling_utils import PreTrainedModel
+from transformers.utils import logging
+from transformers.utils.deprecation import deprecate_kwarg
+
+from fla.layers.abc import ABCAttention
+from fla.layers.attn import Attention
+from fla.models.abc.configuration_abc import ABCConfig
+from fla.models.utils import Cache
+from fla.modules import FusedCrossEntropyLoss, FusedLinearCrossEntropyLoss
+from fla.modules import GatedMLP as ABCMLP
+from fla.modules import RMSNorm
+
+logger = logging.get_logger(__name__)
+
+if TYPE_CHECKING:
+    from transformers.processing_utils import Unpack
+
+
+class ABCBlock(nn.Module):
+    def __init__(self, config: ABCConfig, layer_idx: int):
+        super().__init__()
+
+        self.config = config
+        self.layer_idx = layer_idx
+
+        self.attn_norm = (RMSNorm if config.fuse_norm else nn.RMSNorm)(config.hidden_size, eps=config.norm_eps)
+        if config.attn is not None and layer_idx in config.attn['layers']:
+            self.attn = Attention(
+                hidden_size=config.hidden_size,
+                num_heads=config.attn['num_heads'],
+                num_kv_heads=config.attn['num_kv_heads'],
+                qkv_bias=config.attn['qkv_bias'],
+                window_size=config.attn['window_size'],
+                rope_theta=config.attn['rope_theta'],
+                max_position_embeddings=config.max_position_embeddings,
+                layer_idx=layer_idx
+            )
+        else:
+            self.attn = ABCAttention(
+                hidden_size=config.hidden_size,
+                expand_k=config.expand_k,
+                expand_v=config.expand_v,
+                num_heads=config.num_heads,
+                num_slots=config.num_slots,
+                use_short_conv=config.use_short_conv,
+                conv_size=config.conv_size,
+                gate_fn=config.hidden_act,
+                elementwise_affine=config.elementwise_affine,
+                norm_eps=config.norm_eps,
+                use_rope=config.use_rope,
+                clamp_min=config.clamp_min,
+                clamp_max=config.clamp_max,
+                fuse_norm=config.fuse_norm,
+                layer_idx=layer_idx
+            )
+        self.mlp_norm = (RMSNorm if config.fuse_norm else nn.RMSNorm)(config.hidden_size, eps=config.norm_eps)
+        self.mlp = ABCMLP(
+            hidden_size=config.hidden_size,
+            hidden_ratio=config.hidden_ratio,
+            intermediate_size=config.intermediate_size,
+            hidden_act=config.hidden_act,
+            fuse_swiglu=config.fuse_swiglu
+        )
+
+    def forward(
+        self,
+        hidden_states: torch.Tensor,
+        attention_mask: Optional[torch.Tensor] = None,
+        past_key_values: Optional[Union[Cache, List[torch.FloatTensor]]] = None,
+        use_cache: Optional[bool] = False,
+        output_attentions: Optional[bool] = False,
+        **kwargs: Unpack[Dict]
+    ) -> Tuple[torch.FloatTensor, Optional[Tuple[torch.FloatTensor, torch.FloatTensor]]]:
+
+        residual = hidden_states
+
+        hidden_states = self.attn_norm(hidden_states)
+        hidden_states, attentions, past_key_values = self.attn(
+            hidden_states=hidden_states,
+            attention_mask=attention_mask,
+            past_key_values=past_key_values,
+            use_cache=use_cache,
+            output_attentions=output_attentions,
+            **kwargs
+        )
+        if self.config.fuse_norm:
+            hidden_states, residual = self.mlp_norm(hidden_states, residual, True)
+        else:
+            hidden_states = residual + hidden_states
+            residual = hidden_states
+            hidden_states = self.mlp_norm(hidden_states)
+        hidden_states = self.mlp(hidden_states)
+        hidden_states = residual + hidden_states
+
+        outputs = (hidden_states, attentions, past_key_values)
+
+        return outputs
+
+
+class ABCPreTrainedModel(PreTrainedModel):
+
+    config_class = ABCConfig
+    base_model_prefix = 'model'
+    supports_gradient_checkpointing = True
+    _no_split_modules = ['ABCBlock']
+    _supports_cache_class = True
+
+    def __init__(self, *inputs, **kwargs):
+        super().__init__(*inputs, **kwargs)
+
+    def _init_weights(
+        self,
+        module: nn.Module,
+        prenorm_residual_strategy: Optional[str] = 'rescale',
+        num_residuals_per_layer: int = 2,
+    ):
+        if isinstance(module, (nn.Linear, nn.Conv1d)):
+            # Slightly different from the TF version which uses truncated_normal for initialization
+            # cf https://github.com/pytorch/pytorch/pull/5617
+            nn.init.normal_(module.weight, mean=0.0, std=self.config.initializer_range)
+            if module.bias is not None:
+                nn.init.zeros_(module.bias)
+        elif isinstance(module, nn.Embedding):
+            nn.init.normal_(module.weight, mean=0.0, std=self.config.initializer_range)
+        elif hasattr(module, 'reset_parameters'):
+            module.reset_parameters()
+
+        if prenorm_residual_strategy is not None:
+            # Reinitialize selected weights subject to the OpenAI GPT-2 Paper Scheme:
+            #   > A modified initialization which accounts for the accumulation on the residual path with model depth. Scale
+            #   > the weights of residual layers at initialization by a factor of 1/√N where N is the # of residual layers.
+            #   >   -- GPT-2 :: https://openai.com/blog/better-language-models/
+            #
+            # Reference (Megatron-LM): https://github.com/NVIDIA/Megatron-LM/blob/main/megatron/model/gpt_model.py
+            p = None
+            if hasattr(module, 'o_proj'):
+                p = module.o_proj.weight
+            elif hasattr(module, 'down_proj'):
+                p = module.down_proj.weight
+            if p is not None:
+                # Special Scaled Initialization --> There are 2 Layer Norms per Transformer Block
+                # Following Pytorch init, except scale by 1/sqrt(2 * n_layer)
+                # We need to reinit p since this code could be called multiple times
+                # Having just p *= scale would repeatedly scale it down
+                if prenorm_residual_strategy == 'rescale':
+                    nn.init.kaiming_uniform_(p, a=math.sqrt(5))
+                    with torch.no_grad():
+                        p /= math.sqrt(num_residuals_per_layer * self.config.num_hidden_layers)
+                elif prenorm_residual_strategy == 'zero':
+                    nn.init.zeros_(p)
+                else:
+                    raise ValueError(f"Invalid prenorm_residual_strategy: {prenorm_residual_strategy}")
+
+
+class ABCModel(ABCPreTrainedModel):
+
+    def __init__(self, config: ABCConfig):
+        super().__init__(config)
+        self.padding_idx = config.pad_token_id
+        self.vocab_size = config.vocab_size
+
+        self.embeddings = nn.Embedding(config.vocab_size, config.hidden_size, self.padding_idx)
+        self.layers = nn.ModuleList([ABCBlock(config, layer_idx) for layer_idx in range(config.num_hidden_layers)])
+        self.norm = (RMSNorm if config.fuse_norm else nn.RMSNorm)(config.hidden_size, eps=config.norm_eps)
+
+        self.gradient_checkpointing = False
+
+        self.post_init()
+
+    def get_input_embeddings(self):
+        return self.embeddings
+
+    def set_input_embeddings(self, value):
+        self.embeddings = value
+
+    def forward(
+        self,
+        input_ids: Optional[torch.LongTensor] = None,
+        attention_mask: Optional[torch.Tensor] = None,  # noqa
+        inputs_embeds: Optional[torch.FloatTensor] = None,
+        past_key_values: Optional[Union[Cache, List[torch.FloatTensor]]] = None,
+        use_cache: Optional[bool] = None,
+        output_attentions: Optional[bool] = None,
+        output_hidden_states: Optional[bool] = None,
+        return_dict: Optional[bool] = None,
+        **kwargs: Unpack[Dict]
+    ) -> Union[Tuple, BaseModelOutputWithPast]:
+        if output_attentions:
+            warnings.warn("`ABCModel` does not `output_attentions` now, setting it to `False`.")
+            output_attentions = False
+        output_attentions = output_attentions if output_attentions is not None else self.config.output_attentions
+        output_hidden_states = output_hidden_states if output_hidden_states is not None else self.config.output_hidden_states
+        use_cache = use_cache if use_cache is not None else (self.config.use_cache if not self.training else False)
+        return_dict = return_dict if return_dict is not None else self.config.use_return_dict
+
+        # retrieve input_ids and inputs_embeds
+        if input_ids is not None and inputs_embeds is not None:
+            raise ValueError("You cannot specify both input_ids and inputs_embeds at the same time")
+        if input_ids is None and inputs_embeds is None:
+            raise ValueError("You have to specify either input_ids or inputs_embeds")
+
+        if inputs_embeds is None:
+            inputs_embeds = self.embeddings(input_ids)
+        hidden_states = inputs_embeds
+
+        if use_cache and not isinstance(past_key_values, Cache):
+            past_key_values = Cache.from_legacy_cache(past_key_values)
+
+        if self.gradient_checkpointing and self.training and use_cache:
+            logger.warning_once("`use_cache=True` is incompatible with gradient checkpointing. Setting `use_cache=False`...")
+            use_cache = False
+
+        all_hidden_states = () if output_hidden_states else None
+        all_attns = () if output_attentions else None
+        for layer in self.layers:
+            if output_hidden_states:
+                all_hidden_states += (hidden_states,)
+
+            if self.gradient_checkpointing and self.training:
+                hidden_states, attentions, past_key_values = self._gradient_checkpointing_func(
+                    layer.__call__,
+                    hidden_states,
+                    attention_mask,
+                    past_key_values,
+                    use_cache,
+                    output_attentions,
+                    **kwargs
+                )
+            else:
+                hidden_states, attentions, past_key_values = layer(
+                    hidden_states,
+                    attention_mask,
+                    past_key_values=past_key_values,
+                    use_cache=use_cache,
+                    output_attentions=output_attentions,
+                    **kwargs
+                )
+
+            if output_attentions:
+                all_attns += (attentions,)
+
+        hidden_states = self.norm(hidden_states)
+
+        # add hidden states from the last decoder layer
+        if output_hidden_states:
+            all_hidden_states += (hidden_states,)
+
+        if not return_dict:
+            return tuple(i for i in [hidden_states, past_key_values, all_hidden_states, all_attns] if i is not None)
+        return BaseModelOutputWithPast(
+            last_hidden_state=hidden_states,
+            past_key_values=past_key_values,
+            hidden_states=all_hidden_states,
+            attentions=all_attns
+        )
+
+
+class ABCForCausalLM(ABCPreTrainedModel, GenerationMixin):
+
+    _tied_weights_keys = ["lm_head.weight"]
+
+    def __init__(self, config):
+        super().__init__(config)
+        self.model = ABCModel(config)
+        self.vocab_size = config.vocab_size
+        self.lm_head = nn.Linear(config.hidden_size, config.vocab_size, bias=False)
+        self.criterion = None
+
+        # Initialize weights and apply final processing
+        self.post_init()
+
+    def get_input_embeddings(self):
+        return self.model.embeddings
+
+    def set_input_embeddings(self, value):
+        self.model.embeddings = value
+
+    def get_output_embeddings(self):
+        return self.lm_head
+
+    def set_output_embeddings(self, new_embeddings):
+        self.lm_head = new_embeddings
+
+    def set_decoder(self, decoder):
+        self.model = decoder
+
+    def get_decoder(self):
+        return self.model
+
+    def generate(self, *args, **kwargs):
+        try:
+            return super().generate(*args, **kwargs)
+        except AttributeError as exception:
+            if 'past_key_values' in str(exception):
+                raise AttributeError(
+                    f"You tried to call `generate` with a decoding strategy that manipulates `past_key_values`, "
+                    f"which is not supported for {self.__class__.__name__}. "
+                    f"Try another generation strategy instead. "
+                    f"For the available generation strategies, check this doc: "
+                    f"https://huggingface.co/docs/transformers/en/generation_strategies#decoding-strategies"
+                )
+            else:
+                raise exception
+
+    @deprecate_kwarg("num_logits_to_keep", version="4.50", new_name="logits_to_keep")
+    def prepare_inputs_for_generation(
+        self,
+        input_ids: torch.LongTensor = None,
+        past_key_values: Optional[Union[Cache, List[torch.FloatTensor]]] = None,
+        attention_mask: Optional[torch.Tensor] = None,
+        inputs_embeds: Optional[torch.Tensor] = None,
+        use_cache: bool = True,
+        logits_to_keep: Optional[int] = None,
+        **kwargs
+    ):
+        # only last token for `inputs_ids` if the `past_key_values` is not empty.
+        if past_key_values is not None and len(past_key_values) > 0:
+            input_ids = input_ids[:, -1:]
+        # if `inputs_embeds` are passed, we only want to use them in the 1st generation step
+        if inputs_embeds is not None and len(past_key_values) == 0:
+            model_inputs = {'inputs_embeds': inputs_embeds}
+        else:
+            # The `contiguous()` here is necessary to have a static stride during decoding. torchdynamo otherwise
+            # recompiles graphs as the stride of the inputs is a guard.
+            # Ref: https://github.com/huggingface/transformers/pull/29114
+            # TODO: use `next_tokens` directly instead.
+            model_inputs = {'input_ids': input_ids.contiguous()}
+
+        if logits_to_keep is not None:
+            model_inputs['logits_to_keep'] = logits_to_keep
+
+        model_inputs.update({
+            'past_key_values': past_key_values,
+            'use_cache': use_cache,
+            'attention_mask': attention_mask,
+        })
+        return model_inputs
+
+    @deprecate_kwarg("num_logits_to_keep", version="4.50", new_name="logits_to_keep")
+    def forward(
+        self,
+        input_ids: torch.LongTensor = None,
+        attention_mask: Optional[torch.Tensor] = None,
+        inputs_embeds: Optional[torch.Tensor] = None,
+        past_key_values: Optional[Union[Cache, List[torch.FloatTensor]]] = None,
+        labels: Optional[torch.LongTensor] = None,
+        use_cache: Optional[bool] = None,
+        output_attentions: Optional[bool] = None,
+        output_hidden_states: Optional[bool] = None,
+        return_dict: Optional[bool] = None,
+        logits_to_keep: Optional[int] = 0,
+        **kwargs: Unpack[Dict]
+    ) -> Union[Tuple, CausalLMOutputWithPast]:
+        output_attentions = output_attentions if output_attentions is not None else self.config.output_attentions
+        output_hidden_states = (
+            output_hidden_states if output_hidden_states is not None else self.config.output_hidden_states
+        )
+        return_dict = return_dict if return_dict is not None else self.config.use_return_dict
+
+        outputs = self.model(
+            input_ids=input_ids,
+            attention_mask=attention_mask,
+            inputs_embeds=inputs_embeds,
+            past_key_values=past_key_values,
+            use_cache=use_cache,
+            output_attentions=output_attentions,
+            output_hidden_states=output_hidden_states,
+            return_dict=return_dict,
+            **kwargs
+        )
+
+        hidden_states = outputs[0]
+        fuse_linear_and_cross_entropy = self.config.fuse_cross_entropy and self.training
+
+        loss, logits = None, None
+        if not fuse_linear_and_cross_entropy or labels is None:
+            logits = self.lm_head(hidden_states if logits_to_keep is None else hidden_states[:, -logits_to_keep:])
+        if labels is not None:
+            if getattr(self, 'criterion', None) is None:
+                if fuse_linear_and_cross_entropy:
+                    criterion = FusedLinearCrossEntropyLoss()
+                elif self.config.fuse_cross_entropy:
+                    criterion = FusedCrossEntropyLoss(inplace_backward=True)
+                else:
+                    criterion = nn.CrossEntropyLoss()
+            else:
+                criterion = self.criterion
+            labels = labels.to(hidden_states.device)
+            labels = torch.cat((labels[..., 1:], torch.full_like(labels[:, :1], criterion.ignore_index)), 1)
+            if fuse_linear_and_cross_entropy:
+                loss = criterion(hidden_states, labels, self.lm_head.weight, self.lm_head.bias)
+            else:
+                loss = criterion(logits.view(labels.numel(), -1), labels.view(-1))
+
+        if not return_dict:
+            output = (logits,) + outputs[1:]
+            return (loss,) + output if loss is not None else output
+
+        return CausalLMOutputWithPast(
+            loss=loss,
+            logits=logits,
+            past_key_values=outputs.past_key_values,
+            hidden_states=outputs.hidden_states,
+            attentions=outputs.attentions,
+        )

fla/models/bitnet/__init__.py

@@ -0,0 +1,13 @@
+# -*- coding: utf-8 -*-
+
+from transformers import AutoConfig, AutoModel, AutoModelForCausalLM
+
+from fla.models.bitnet.configuration_bitnet import BitNetConfig
+from fla.models.bitnet.modeling_bitnet import BitNetForCausalLM, BitNetModel
+
+AutoConfig.register(BitNetConfig.model_type, BitNetConfig)
+AutoModel.register(BitNetConfig, BitNetModel)
+AutoModelForCausalLM.register(BitNetConfig, BitNetForCausalLM)
+
+
+__all__ = ['BitNetConfig', 'BitNetForCausalLM', 'BitNetModel']

fla/models/forgetting_transformer/__init__.py

@@ -0,0 +1,16 @@
+# -*- coding: utf-8 -*-
+
+from transformers import AutoConfig, AutoModel, AutoModelForCausalLM
+
+from fla.models.forgetting_transformer.configuration_forgetting_transformer import ForgettingTransformerConfig
+from fla.models.forgetting_transformer.modeling_forgetting_transformer import (
+    ForgettingTransformerForCausalLM,
+    ForgettingTransformerModel
+)
+
+AutoConfig.register(ForgettingTransformerConfig.model_type, ForgettingTransformerConfig)
+AutoModel.register(ForgettingTransformerConfig, ForgettingTransformerModel)
+AutoModelForCausalLM.register(ForgettingTransformerConfig, ForgettingTransformerForCausalLM)
+
+
+__all__ = ['ForgettingTransformerConfig', 'ForgettingTransformerForCausalLM', 'ForgettingTransformerModel']

fla/models/forgetting_transformer/configuration_forgetting_transformer.py

@@ -0,0 +1,68 @@
+# -*- coding: utf-8 -*-
+
+from typing import Optional
+
+from transformers.configuration_utils import PretrainedConfig
+
+
+class ForgettingTransformerConfig(PretrainedConfig):
+
+    model_type = 'forgetting_transformer'
+    keys_to_ignore_at_inference = ['past_key_values']
+
+    def __init__(
+        self,
+        hidden_size: int = 2048,
+        num_hidden_layers: int = 24,
+        num_heads: int = 32,
+        num_kv_heads: Optional[int] = None,
+        qkv_bias: bool = False,
+        qk_norm: bool = False,
+        window_size: Optional[int] = None,
+        use_output_gate: bool = False,
+        hidden_ratio: Optional[int] = 4,
+        intermediate_size: Optional[int] = None,
+        hidden_act: str = "swish",
+        initializer_range: float = 0.006,
+        elementwise_affine: Optional[bool] = True,
+        norm_eps: float = 1e-6,
+        use_cache: bool = True,
+        pad_token_id: Optional[int] = None,
+        bos_token_id: int = 1,
+        eos_token_id: int = 2,
+        tie_word_embeddings: bool = False,
+        fuse_norm: bool = True,
+        fuse_swiglu: bool = True,
+        fuse_cross_entropy: bool = True,
+        vocab_size: int = 32000,
+        **kwargs,
+    ):
+        self.hidden_size = hidden_size
+        self.num_hidden_layers = num_hidden_layers
+        self.num_heads = num_heads
+        self.num_kv_heads = num_kv_heads
+        self.qkv_bias = qkv_bias
+        self.qk_norm = qk_norm
+        self.window_size = window_size
+        self.use_output_gate = use_output_gate
+        self.hidden_ratio = hidden_ratio
+        self.intermediate_size = intermediate_size
+        self.hidden_act = hidden_act
+
+        self.initializer_range = initializer_range
+        self.elementwise_affine = elementwise_affine
+        self.norm_eps = norm_eps
+        self.use_cache = use_cache
+
+        self.fuse_norm = fuse_norm
+        self.fuse_swiglu = fuse_swiglu
+        self.fuse_cross_entropy = fuse_cross_entropy
+        self.vocab_size = vocab_size
+
+        super().__init__(
+            pad_token_id=pad_token_id,
+            bos_token_id=bos_token_id,
+            eos_token_id=eos_token_id,
+            tie_word_embeddings=tie_word_embeddings,
+            **kwargs,
+        )

fla/models/gated_deltanet/__init__.py

@@ -0,0 +1,12 @@
+# -*- coding: utf-8 -*-
+
+from transformers import AutoConfig, AutoModel, AutoModelForCausalLM
+
+from fla.models.gated_deltanet.configuration_gated_deltanet import GatedDeltaNetConfig
+from fla.models.gated_deltanet.modeling_gated_deltanet import GatedDeltaNetForCausalLM, GatedDeltaNetModel
+
+AutoConfig.register(GatedDeltaNetConfig.model_type, GatedDeltaNetConfig)
+AutoModel.register(GatedDeltaNetConfig, GatedDeltaNetModel)
+AutoModelForCausalLM.register(GatedDeltaNetConfig, GatedDeltaNetForCausalLM)
+
+__all__ = ['GatedDeltaNetConfig', 'GatedDeltaNetForCausalLM', 'GatedDeltaNetModel']

fla/models/gated_deltanet/configuration_gated_deltanet.py

@@ -0,0 +1,83 @@
+# -*- coding: utf-8 -*-
+
+from typing import Dict, Optional
+
+from transformers.configuration_utils import PretrainedConfig
+
+
+class GatedDeltaNetConfig(PretrainedConfig):
+    model_type = 'gated_deltanet'
+    keys_to_ignore_at_inference = ['past_key_values']
+
+    def __init__(
+        self,
+        attn_mode: str = "chunk",
+        hidden_size: int = 2048,
+        expand_v: int = 2,
+        use_gate: bool = True,
+        use_short_conv: bool = True,
+        conv_size: int = 4,
+        head_dim: int = 256,
+        num_heads: int = 6,
+        max_position_embeddings: int = 2048,
+        hidden_ratio: Optional[int] = 4,
+        intermediate_size: Optional[int] = None,
+        hidden_act: str = "swish",
+        num_hidden_layers: int = 21,
+        norm_eps: float = 1e-6,
+        attn: Optional[Dict] = None,
+        use_cache: bool = True,
+        pad_token_id: int = None,
+        bos_token_id: int = 1,
+        eos_token_id: int = 2,
+        tie_word_embeddings: bool = False,
+        initializer_range: float = 0.006,
+        fuse_norm: bool = True,
+        fuse_swiglu: bool = True,
+        fuse_cross_entropy: bool = True,
+        vocab_size: int = 32000,
+        **kwargs
+    ):
+        self.attn_mode = attn_mode
+        self.hidden_size = hidden_size
+        self.expand_v = expand_v
+        self.use_gate = use_gate
+        self.use_short_conv = use_short_conv
+        self.conv_size = conv_size
+        self.head_dim = head_dim
+        self.num_heads = num_heads
+        self.max_position_embeddings = max_position_embeddings
+
+        self.hidden_ratio = hidden_ratio
+        self.intermediate_size = intermediate_size
+        self.hidden_act = hidden_act
+        self.num_hidden_layers = num_hidden_layers
+        self.norm_eps = norm_eps
+        self.attn = attn
+        self.use_cache = use_cache
+        self.initializer_range = initializer_range
+
+        self.fuse_norm = fuse_norm
+        self.fuse_swiglu = fuse_swiglu
+        self.fuse_cross_entropy = fuse_cross_entropy
+        self.vocab_size = vocab_size
+
+        if attn is not None:
+            if not isinstance(attn, Dict):
+                raise ValueError("attn must be a dictionary")
+            if 'layers' not in attn:
+                raise ValueError("Layer indices must be provided to initialize hybrid attention layers")
+            if 'num_heads' not in attn:
+                raise ValueError("Number of heads must be provided to initialize hybrid attention layers")
+            attn['num_kv_heads'] = attn.get('num_kv_heads', attn['num_heads'])
+            attn['qkv_bias'] = attn.get('qkv_bias', False)
+            attn['window_size'] = attn.get('window_size', None)
+            attn['rope_theta'] = attn.get('rope_theta', 10000.)
+
+        super().__init__(
+            pad_token_id=pad_token_id,
+            bos_token_id=bos_token_id,
+            eos_token_id=eos_token_id,
+            tie_word_embeddings=tie_word_embeddings,
+            **kwargs,
+        )

fla/models/gated_deltaproduct/__init__.py

@@ -0,0 +1,14 @@
+from transformers import AutoConfig, AutoModel, AutoModelForCausalLM
+
+from fla.models.gated_deltaproduct.configuration_gated_deltaproduct import GatedDeltaProductConfig
+from fla.models.gated_deltaproduct.modeling_gated_deltaproduct import GatedDeltaProductForCausalLM, GatedDeltaProductModel
+
+AutoConfig.register(GatedDeltaProductConfig.model_type, GatedDeltaProductConfig)
+AutoModel.register(GatedDeltaProductConfig, GatedDeltaProductModel)
+AutoModelForCausalLM.register(GatedDeltaProductConfig, GatedDeltaProductForCausalLM)
+
+__all__ = [
+    "GatedDeltaProductConfig",
+    "GatedDeltaProductForCausalLM",
+    "GatedDeltaProductModel",
+]

fla/models/gla/__init__.py

@@ -0,0 +1,13 @@
+# -*- coding: utf-8 -*-
+
+from transformers import AutoConfig, AutoModel, AutoModelForCausalLM
+
+from fla.models.gla.configuration_gla import GLAConfig
+from fla.models.gla.modeling_gla import GLAForCausalLM, GLAModel
+
+AutoConfig.register(GLAConfig.model_type, GLAConfig)
+AutoModel.register(GLAConfig, GLAModel)
+AutoModelForCausalLM.register(GLAConfig, GLAForCausalLM)
+
+
+__all__ = ['GLAConfig', 'GLAForCausalLM', 'GLAModel']

fla/models/gla/configuration_gla.py

@@ -0,0 +1,95 @@
+# -*- coding: utf-8 -*-
+
+from typing import Dict, Optional
+
+from transformers.configuration_utils import PretrainedConfig
+
+
+class GLAConfig(PretrainedConfig):
+
+    model_type = 'gla'
+    keys_to_ignore_at_inference = ['past_key_values']
+
+    def __init__(
+        self,
+        hidden_size: int = 2048,
+        expand_k: int = 0.5,
+        expand_v: int = 1,
+        hidden_ratio: Optional[int] = 4,
+        intermediate_size: Optional[int] = None,
+        num_hidden_layers: int = 24,
+        num_heads: int = 4,
+        num_kv_heads: Optional[int] = None,
+        feature_map: Optional[str] = None,
+        attn_mode: str = "chunk",
+        use_short_conv: bool = False,
+        conv_size: int = 4,
+        use_output_gate: bool = True,
+        clamp_min: Optional[float] = None,
+        hidden_act: str = "swish",
+        max_position_embeddings: int = 2048,
+        elementwise_affine: Optional[bool] = True,
+        norm_eps: float = 1e-6,
+        use_gk: bool = True,
+        use_gv: bool = False,
+        attn: Optional[Dict] = None,
+        use_cache: bool = True,
+        pad_token_id: int = None,
+        bos_token_id: int = 1,
+        eos_token_id: int = 2,
+        tie_word_embeddings: bool = False,
+        initializer_range: float = 0.006,
+        fuse_norm: bool = True,
+        fuse_swiglu: bool = True,
+        fuse_cross_entropy: bool = True,
+        vocab_size: int = 32000,
+        **kwargs
+    ):
+        self.hidden_size = hidden_size
+        self.expand_k = expand_k
+        self.expand_v = expand_v
+        self.hidden_ratio = hidden_ratio
+        self.intermediate_size = intermediate_size
+        self.num_hidden_layers = num_hidden_layers
+        self.num_heads = num_heads
+        self.num_kv_heads = num_kv_heads
+        self.feature_map = feature_map
+        self.attn_mode = attn_mode
+        self.use_short_conv = use_short_conv
+        self.conv_size = conv_size
+        self.use_output_gate = use_output_gate
+        self.clamp_min = clamp_min
+        self.hidden_act = hidden_act
+        self.max_position_embeddings = max_position_embeddings
+        self.elementwise_affine = elementwise_affine
+        self.norm_eps = norm_eps
+        self.use_gk = use_gk
+        self.use_gv = use_gv
+        self.attn = attn
+        self.use_cache = use_cache
+        self.initializer_range = initializer_range
+
+        self.fuse_norm = fuse_norm
+        self.fuse_swiglu = fuse_swiglu
+        self.fuse_cross_entropy = fuse_cross_entropy
+        self.vocab_size = vocab_size
+
+        if attn is not None:
+            if not isinstance(attn, Dict):
+                raise ValueError("attn must be a dictionary")
+            if 'layers' not in attn:
+                raise ValueError("Layer indices must be provided to initialize hybrid attention layers")
+            if 'num_heads' not in attn:
+                raise ValueError("Number of heads must be provided to initialize hybrid attention layers")
+            attn['num_kv_heads'] = attn.get('num_kv_heads', attn['num_heads'])
+            attn['qkv_bias'] = attn.get('qkv_bias', False)
+            attn['window_size'] = attn.get('window_size', None)
+            attn['rope_theta'] = attn.get('rope_theta', 10000.)
+
+        super().__init__(
+            pad_token_id=pad_token_id,
+            bos_token_id=bos_token_id,
+            eos_token_id=eos_token_id,
+            tie_word_embeddings=tie_word_embeddings,
+            **kwargs,
+        )

fla/models/gla/modeling_gla.py

@@ -0,0 +1,417 @@
+# -*- coding: utf-8 -*-
+
+from __future__ import annotations
+
+import math
+import warnings
+from typing import TYPE_CHECKING, Dict, List, Optional, Tuple, Union
+
+import torch
+import torch.nn as nn
+import torch.utils.checkpoint
+from transformers.generation import GenerationMixin
+from transformers.modeling_outputs import BaseModelOutputWithPast, CausalLMOutputWithPast
+from transformers.modeling_utils import PreTrainedModel
+from transformers.utils import logging
+from transformers.utils.deprecation import deprecate_kwarg
+
+from fla.layers.attn import Attention
+from fla.layers.gla import GatedLinearAttention
+from fla.models.gla.configuration_gla import GLAConfig
+from fla.models.utils import Cache
+from fla.modules import FusedCrossEntropyLoss, FusedLinearCrossEntropyLoss
+from fla.modules import GatedMLP as GLAMLP
+from fla.modules import RMSNorm
+
+if TYPE_CHECKING:
+    from transformers.processing_utils import Unpack
+
+logger = logging.get_logger(__name__)
+
+
+class GLABlock(nn.Module):
+    def __init__(self, config: GLAConfig, layer_idx: int):
+        super().__init__()
+
+        self.config = config
+        self.layer_idx = layer_idx
+
+        self.attn_norm = (RMSNorm if config.fuse_norm else nn.RMSNorm)(config.hidden_size, eps=config.norm_eps)
+        if config.attn is not None and layer_idx in config.attn['layers']:
+            self.attn = Attention(
+                hidden_size=config.hidden_size,
+                num_heads=config.attn['num_heads'],
+                num_kv_heads=config.attn['num_kv_heads'],
+                qkv_bias=config.attn['qkv_bias'],
+                window_size=config.attn['window_size'],
+                rope_theta=config.attn['rope_theta'],
+                max_position_embeddings=config.max_position_embeddings,
+                layer_idx=layer_idx
+            )
+        else:
+            self.attn = GatedLinearAttention(
+                mode=config.attn_mode,
+                hidden_size=config.hidden_size,
+                expand_k=config.expand_k,
+                expand_v=config.expand_v,
+                num_heads=config.num_heads,
+                num_kv_heads=config.num_kv_heads,
+                feature_map=config.feature_map,
+                use_short_conv=config.use_short_conv,
+                conv_size=config.conv_size,
+                use_output_gate=config.use_output_gate,
+                gate_fn=config.hidden_act,
+                elementwise_affine=config.elementwise_affine,
+                norm_eps=config.norm_eps,
+                clamp_min=config.clamp_min,
+                fuse_norm=config.fuse_norm,
+                layer_idx=layer_idx
+            )
+        self.mlp_norm = (RMSNorm if config.fuse_norm else nn.RMSNorm)(config.hidden_size, eps=config.norm_eps)
+        self.mlp = GLAMLP(
+            hidden_size=config.hidden_size,
+            hidden_ratio=config.hidden_ratio,
+            intermediate_size=config.intermediate_size,
+            hidden_act=config.hidden_act,
+            fuse_swiglu=config.fuse_swiglu
+        )
+
+    def forward(
+        self,
+        hidden_states: torch.Tensor,
+        attention_mask: Optional[torch.Tensor] = None,
+        past_key_values: Optional[Union[Cache, List[torch.FloatTensor]]] = None,
+        use_cache: Optional[bool] = False,
+        output_attentions: Optional[bool] = False,
+        **kwargs: Unpack[Dict]
+    ) -> Tuple[torch.FloatTensor, Optional[Tuple[torch.FloatTensor, torch.FloatTensor]]]:
+        residual = hidden_states
+        hidden_states = self.attn_norm(hidden_states)
+        hidden_states, attentions, past_key_values = self.attn(
+            hidden_states=hidden_states,
+            attention_mask=attention_mask,
+            past_key_values=past_key_values,
+            use_cache=use_cache,
+            output_attentions=output_attentions,
+            **kwargs
+        )
+        if self.config.fuse_norm:
+            hidden_states, residual = self.mlp_norm(hidden_states, residual, True)
+        else:
+            hidden_states = residual + hidden_states
+            residual = hidden_states
+            hidden_states = self.mlp_norm(hidden_states)
+        hidden_states = self.mlp(hidden_states, **kwargs)
+        hidden_states = residual + hidden_states
+
+        outputs = (hidden_states, attentions, past_key_values)
+
+        return outputs
+
+
+class GLAPreTrainedModel(PreTrainedModel):
+
+    config_class = GLAConfig
+    base_model_prefix = 'model'
+    supports_gradient_checkpointing = True
+    _no_split_modules = ['GLABlock']
+    _supports_cache_class = True
+
+    def __init__(self, *inputs, **kwargs):
+        super().__init__(*inputs, **kwargs)
+
+    def _init_weights(
+        self,
+        module: nn.Module,
+        prenorm_residual_strategy: Optional[str] = 'rescale',
+        num_residuals_per_layer: int = 2,
+    ):
+        if isinstance(module, (nn.Linear, nn.Conv1d)):
+            # Slightly different from the TF version which uses truncated_normal for initialization
+            # cf https://github.com/pytorch/pytorch/pull/5617
+            nn.init.normal_(module.weight, mean=0.0, std=self.config.initializer_range)
+            if module.bias is not None:
+                nn.init.zeros_(module.bias)
+        elif isinstance(module, nn.Embedding):
+            nn.init.normal_(module.weight, mean=0.0, std=self.config.initializer_range)
+        elif hasattr(module, 'reset_parameters'):
+            module.reset_parameters()
+
+        if prenorm_residual_strategy is not None:
+            # Reinitialize selected weights subject to the OpenAI GPT-2 Paper Scheme:
+            #   > A modified initialization which accounts for the accumulation on the residual path with model depth. Scale
+            #   > the weights of residual layers at initialization by a factor of 1/√N where N is the # of residual layers.
+            #   >   -- GPT-2 :: https://openai.com/blog/better-language-models/
+            #
+            # Reference (Megatron-LM): https://github.com/NVIDIA/Megatron-LM/blob/main/megatron/model/gpt_model.py
+            p = None
+            if hasattr(module, 'o_proj'):
+                p = module.o_proj.weight
+            elif hasattr(module, 'down_proj'):
+                p = module.down_proj.weight
+            if p is not None:
+                # Special Scaled Initialization --> There are 2 Layer Norms per Transformer Block
+                # Following Pytorch init, except scale by 1/sqrt(2 * n_layer)
+                # We need to reinit p since this code could be called multiple times
+                # Having just p *= scale would repeatedly scale it down
+                if prenorm_residual_strategy == 'rescale':
+                    nn.init.kaiming_uniform_(p, a=math.sqrt(5))
+                    with torch.no_grad():
+                        p /= math.sqrt(num_residuals_per_layer * self.config.num_hidden_layers)
+                elif prenorm_residual_strategy == 'zero':
+                    nn.init.zeros_(p)
+                else:
+                    raise ValueError(f"Invalid prenorm_residual_strategy: {prenorm_residual_strategy}")
+
+
+class GLAModel(GLAPreTrainedModel):
+
+    def __init__(self, config: GLAConfig):
+        super().__init__(config)
+        self.padding_idx = config.pad_token_id
+        self.vocab_size = config.vocab_size
+
+        self.embeddings = nn.Embedding(config.vocab_size, config.hidden_size, self.padding_idx)
+        self.layers = nn.ModuleList([GLABlock(config, layer_idx) for layer_idx in range(config.num_hidden_layers)])
+        self.norm = (RMSNorm if config.fuse_norm else nn.RMSNorm)(config.hidden_size, eps=config.norm_eps)
+
+        self.gradient_checkpointing = False
+
+        self.post_init()
+
+    def get_input_embeddings(self):
+        return self.embeddings
+
+    def set_input_embeddings(self, value):
+        self.embeddings = value
+
+    def forward(
+        self,
+        input_ids: Optional[torch.LongTensor] = None,
+        attention_mask: Optional[torch.Tensor] = None,  # noqa
+        inputs_embeds: Optional[torch.FloatTensor] = None,
+        past_key_values: Optional[Union[Cache, List[torch.FloatTensor]]] = None,
+        use_cache: Optional[bool] = None,
+        output_attentions: Optional[bool] = None,
+        output_hidden_states: Optional[bool] = None,
+        return_dict: Optional[bool] = None,
+        **kwargs: Unpack[Dict]
+    ) -> Union[Tuple, BaseModelOutputWithPast]:
+        if output_attentions:
+            warnings.warn("`GLAModel` does not `output_attentions` now, setting it to `False`.")
+            output_attentions = False
+        output_attentions = output_attentions if output_attentions is not None else self.config.output_attentions
+        output_hidden_states = output_hidden_states if output_hidden_states is not None else self.config.output_hidden_states
+        use_cache = use_cache if use_cache is not None else (self.config.use_cache if not self.training else False)
+        return_dict = return_dict if return_dict is not None else self.config.use_return_dict
+
+        # retrieve input_ids and inputs_embeds
+        if input_ids is not None and inputs_embeds is not None:
+            raise ValueError("You cannot specify both input_ids and inputs_embeds at the same time")
+        if input_ids is None and inputs_embeds is None:
+            raise ValueError("You have to specify either input_ids or inputs_embeds")
+
+        if inputs_embeds is None:
+            inputs_embeds = self.embeddings(input_ids)
+        hidden_states = inputs_embeds
+
+        if use_cache and not isinstance(past_key_values, Cache):
+            past_key_values = Cache.from_legacy_cache(past_key_values)
+
+        if self.gradient_checkpointing and self.training and use_cache:
+            logger.warning_once("`use_cache=True` is incompatible with gradient checkpointing. Setting `use_cache=False`...")
+            use_cache = False
+
+        all_hidden_states = () if output_hidden_states else None
+        all_attns = () if output_attentions else None
+        for layer in self.layers:
+            if output_hidden_states:
+                all_hidden_states += (hidden_states,)
+
+            if self.gradient_checkpointing and self.training:
+                hidden_states, attentions, past_key_values = self._gradient_checkpointing_func(
+                    layer.__call__,
+                    hidden_states,
+                    attention_mask,
+                    past_key_values,
+                    use_cache,
+                    output_attentions,
+                    **kwargs
+                )
+            else:
+                hidden_states, attentions, past_key_values = layer(
+                    hidden_states,
+                    attention_mask=attention_mask,
+                    past_key_values=past_key_values,
+                    use_cache=use_cache,
+                    output_attentions=output_attentions,
+                    **kwargs
+                )
+
+            if output_attentions:
+                all_attns += (attentions,)
+
+        hidden_states = self.norm(hidden_states)
+
+        # add hidden states from the last decoder layer
+        if output_hidden_states:
+            all_hidden_states += (hidden_states,)
+
+        if not return_dict:
+            return tuple(i for i in [hidden_states, past_key_values, all_hidden_states, all_attns] if i is not None)
+        return BaseModelOutputWithPast(
+            last_hidden_state=hidden_states,
+            past_key_values=past_key_values,
+            hidden_states=all_hidden_states,
+            attentions=all_attns
+        )
+
+
+class GLAForCausalLM(GLAPreTrainedModel, GenerationMixin):
+
+    _tied_weights_keys = ["lm_head.weight"]
+
+    def __init__(self, config):
+        super().__init__(config)
+        self.model = GLAModel(config)
+        self.vocab_size = config.vocab_size
+        self.lm_head = nn.Linear(config.hidden_size, config.vocab_size, bias=False)
+        self.criterion = None
+
+        # Initialize weights and apply final processing
+        self.post_init()
+
+    def get_input_embeddings(self):
+        return self.model.embeddings
+
+    def set_input_embeddings(self, value):
+        self.model.embeddings = value
+
+    def get_output_embeddings(self):
+        return self.lm_head
+
+    def set_output_embeddings(self, new_embeddings):
+        self.lm_head = new_embeddings
+
+    def set_decoder(self, decoder):
+        self.model = decoder
+
+    def get_decoder(self):
+        return self.model
+
+    def generate(self, *args, **kwargs):
+        try:
+            return super().generate(*args, **kwargs)
+        except AttributeError as exception:
+            if 'past_key_values' in str(exception):
+                raise AttributeError(
+                    f"You tried to call `generate` with a decoding strategy that manipulates `past_key_values`, "
+                    f"which is not supported for {self.__class__.__name__}. "
+                    f"Try another generation strategy instead. "
+                    f"For the available generation strategies, check this doc: "
+                    f"https://huggingface.co/docs/transformers/en/generation_strategies#decoding-strategies"
+                )
+            else:
+                raise exception
+
+    @deprecate_kwarg("num_logits_to_keep", version="4.50", new_name="logits_to_keep")
+    def prepare_inputs_for_generation(
+        self,
+        input_ids: torch.LongTensor = None,
+        past_key_values: Optional[Union[Cache, List[torch.FloatTensor]]] = None,
+        attention_mask: Optional[torch.Tensor] = None,
+        inputs_embeds: Optional[torch.Tensor] = None,
+        use_cache: bool = True,
+        logits_to_keep: Optional[int] = None,
+        **kwargs
+    ):
+        # only last token for `inputs_ids` if the `past_key_values` is not empty.
+        if past_key_values is not None and len(past_key_values) > 0:
+            input_ids = input_ids[:, -1:]
+        # if `inputs_embeds` are passed, we only want to use them in the 1st generation step
+        if inputs_embeds is not None and len(past_key_values) == 0:
+            model_inputs = {'inputs_embeds': inputs_embeds}
+        else:
+            # The `contiguous()` here is necessary to have a static stride during decoding. torchdynamo otherwise
+            # recompiles graphs as the stride of the inputs is a guard.
+            # Ref: https://github.com/huggingface/transformers/pull/29114
+            # TODO: use `next_tokens` directly instead.
+            model_inputs = {'input_ids': input_ids.contiguous()}
+
+        if logits_to_keep is not None:
+            model_inputs['logits_to_keep'] = logits_to_keep
+
+        model_inputs.update({
+            'past_key_values': past_key_values,
+            'use_cache': use_cache,
+            'attention_mask': attention_mask,
+        })
+        return model_inputs
+
+    @deprecate_kwarg("num_logits_to_keep", version="4.50", new_name="logits_to_keep")
+    def forward(
+        self,
+        input_ids: torch.LongTensor = None,
+        attention_mask: Optional[torch.Tensor] = None,
+        inputs_embeds: Optional[torch.Tensor] = None,
+        past_key_values: Optional[Union[Cache, List[torch.FloatTensor]]] = None,
+        labels: Optional[torch.LongTensor] = None,
+        use_cache: Optional[bool] = None,
+        output_attentions: Optional[bool] = None,
+        output_hidden_states: Optional[bool] = None,
+        return_dict: Optional[bool] = None,
+        logits_to_keep: Optional[int] = 0,
+        **kwargs: Unpack[Dict]
+    ) -> Union[Tuple, CausalLMOutputWithPast]:
+        output_attentions = output_attentions if output_attentions is not None else self.config.output_attentions
+        output_hidden_states = (
+            output_hidden_states if output_hidden_states is not None else self.config.output_hidden_states
+        )
+        return_dict = return_dict if return_dict is not None else self.config.use_return_dict
+
+        outputs = self.model(
+            input_ids=input_ids,
+            attention_mask=attention_mask,
+            inputs_embeds=inputs_embeds,
+            past_key_values=past_key_values,
+            use_cache=use_cache,
+            output_attentions=output_attentions,
+            output_hidden_states=output_hidden_states,
+            return_dict=return_dict,
+            **kwargs
+        )
+
+        hidden_states = outputs[0]
+        fuse_linear_and_cross_entropy = self.config.fuse_cross_entropy and self.training
+
+        loss, logits = None, None
+        if not fuse_linear_and_cross_entropy or labels is None:
+            logits = self.lm_head(hidden_states if logits_to_keep is None else hidden_states[:, -logits_to_keep:])
+        if labels is not None:
+            if getattr(self, 'criterion', None) is None:
+                if fuse_linear_and_cross_entropy:
+                    criterion = FusedLinearCrossEntropyLoss()
+                elif self.config.fuse_cross_entropy:
+                    criterion = FusedCrossEntropyLoss(inplace_backward=True)
+                else:
+                    criterion = nn.CrossEntropyLoss()
+            else:
+                criterion = self.criterion
+            labels = labels.to(hidden_states.device)
+            labels = torch.cat((labels[..., 1:], torch.full_like(labels[:, :1], criterion.ignore_index)), 1)
+            if fuse_linear_and_cross_entropy:
+                loss = criterion(hidden_states, labels, self.lm_head.weight, self.lm_head.bias)
+            else:
+                loss = criterion(logits.view(labels.numel(), -1), labels.view(-1))
+
+        if not return_dict:
+            output = (logits,) + outputs[1:]
+            return (loss,) + output if loss is not None else output
+
+        return CausalLMOutputWithPast(
+            loss=loss,
+            logits=logits,
+            past_key_values=outputs.past_key_values,
+            hidden_states=outputs.hidden_states,
+            attentions=outputs.attentions,
+        )

fla/models/hgrn2/__init__.py

@@ -0,0 +1,13 @@
+# -*- coding: utf-8 -*-
+
+from transformers import AutoConfig, AutoModel, AutoModelForCausalLM
+
+from fla.models.hgrn2.configuration_hgrn2 import HGRN2Config
+from fla.models.hgrn2.modeling_hgrn2 import HGRN2ForCausalLM, HGRN2Model
+
+AutoConfig.register(HGRN2Config.model_type, HGRN2Config)
+AutoModel.register(HGRN2Config, HGRN2Model)
+AutoModelForCausalLM.register(HGRN2Config, HGRN2ForCausalLM)
+
+
+__all__ = ['HGRN2Config', 'HGRN2ForCausalLM', 'HGRN2Model']

fla/models/linear_attn/__init__.py

@@ -0,0 +1,12 @@
+# -*- coding: utf-8 -*-
+
+from transformers import AutoConfig, AutoModel, AutoModelForCausalLM
+
+from fla.models.linear_attn.configuration_linear_attn import LinearAttentionConfig
+from fla.models.linear_attn.modeling_linear_attn import LinearAttentionForCausalLM, LinearAttentionModel
+
+AutoConfig.register(LinearAttentionConfig.model_type, LinearAttentionConfig)
+AutoModel.register(LinearAttentionConfig, LinearAttentionModel)
+AutoModelForCausalLM.register(LinearAttentionConfig, LinearAttentionForCausalLM)
+
+__all__ = ['LinearAttentionConfig', 'LinearAttentionForCausalLM', 'LinearAttentionModel']

fla/models/linear_attn/configuration_linear_attn.py

@@ -0,0 +1,91 @@
+# -*- coding: utf-8 -*-
+
+from typing import Dict, Optional
+
+from transformers.configuration_utils import PretrainedConfig
+
+
+class LinearAttentionConfig(PretrainedConfig):
+
+    model_type = 'linear_attn'
+    keys_to_ignore_at_inference = ['past_key_values']
+
+    def __init__(
+        self,
+        attn_mode: str = "fused_chunk",
+        hidden_size: int = 2048,
+        expand_k: int = 1,
+        expand_v: int = 1,
+        hidden_ratio: Optional[int] = 4,
+        intermediate_size: Optional[int] = None,
+        num_hidden_layers: int = 24,
+        num_heads: int = 4,
+        num_kv_heads: Optional[int] = None,
+        feature_map: str = "elementwise_product",
+        tie_feature_map_qk: bool = False,
+        norm_q: bool = False,
+        norm_k: bool = False,
+        norm_feature_map: bool = False,
+        hidden_act: str = "swish",
+        max_position_embeddings: int = 2048,
+        elementwise_affine: Optional[bool] = True,
+        norm_eps: float = 1e-6,
+        attn: Optional[Dict] = None,
+        use_cache: bool = True,
+        pad_token_id: int = None,
+        bos_token_id: int = 1,
+        eos_token_id: int = 2,
+        tie_word_embeddings: bool = False,
+        initializer_range: float = 0.006,
+        fuse_norm: bool = True,
+        fuse_swiglu: bool = True,
+        fuse_cross_entropy: bool = True,
+        vocab_size: int = 32000,
+        **kwargs
+    ):
+        self.attn_mode = attn_mode
+        self.hidden_size = hidden_size
+        self.expand_k = expand_k
+        self.expand_v = expand_v
+        self.hidden_ratio = hidden_ratio
+        self.intermediate_size = intermediate_size
+        self.num_hidden_layers = num_hidden_layers
+        self.num_heads = num_heads
+        self.num_kv_heads = num_kv_heads
+        self.feature_map = feature_map
+        self.tie_feature_map_qk = tie_feature_map_qk
+        self.norm_q = norm_q
+        self.norm_k = norm_k
+        self.norm_feature_map = norm_feature_map
+        self.hidden_act = hidden_act
+        self.max_position_embeddings = max_position_embeddings
+        self.elementwise_affine = elementwise_affine
+        self.norm_eps = norm_eps
+        self.attn = attn
+        self.use_cache = use_cache
+        self.initializer_range = initializer_range
+
+        self.fuse_norm = fuse_norm
+        self.fuse_swiglu = fuse_swiglu
+        self.fuse_cross_entropy = fuse_cross_entropy
+        self.vocab_size = vocab_size
+
+        if attn is not None:
+            if not isinstance(attn, Dict):
+                raise ValueError("attn must be a dictionary")
+            if 'layers' not in attn:
+                raise ValueError("Layer indices must be provided to initialize hybrid attention layers")
+            if 'num_heads' not in attn:
+                raise ValueError("Number of heads must be provided to initialize hybrid attention layers")
+            attn['num_kv_heads'] = attn.get('num_kv_heads', attn['num_heads'])
+            attn['qkv_bias'] = attn.get('qkv_bias', False)
+            attn['window_size'] = attn.get('window_size', None)
+            attn['rope_theta'] = attn.get('rope_theta', 10000.)
+
+        super().__init__(
+            pad_token_id=pad_token_id,
+            bos_token_id=bos_token_id,
+            eos_token_id=eos_token_id,
+            tie_word_embeddings=tie_word_embeddings,
+            **kwargs,
+        )

fla/models/linear_attn/modeling_linear_attn.py

@@ -0,0 +1,406 @@
+# -*- coding: utf-8 -*-
+
+from __future__ import annotations
+
+import math
+import warnings
+from typing import List, Optional, Tuple, Union
+
+import torch
+import torch.nn as nn
+import torch.utils.checkpoint
+from transformers.generation import GenerationMixin
+from transformers.modeling_outputs import BaseModelOutputWithPast, CausalLMOutputWithPast
+from transformers.modeling_utils import PreTrainedModel
+from transformers.utils import logging
+from transformers.utils.deprecation import deprecate_kwarg
+
+from fla.layers.attn import Attention
+from fla.layers.linear_attn import LinearAttention
+from fla.models.linear_attn.configuration_linear_attn import LinearAttentionConfig
+from fla.models.utils import Cache
+from fla.modules import FusedCrossEntropyLoss, FusedLinearCrossEntropyLoss
+from fla.modules import GatedMLP as LinearAttentionMLP
+from fla.modules import RMSNorm
+
+logger = logging.get_logger(__name__)
+
+
+class LinearAttentionBlock(nn.Module):
+    def __init__(self, config: LinearAttentionConfig, layer_idx: int):
+        super().__init__()
+
+        self.config = config
+        self.layer_idx = layer_idx
+
+        self.attn_norm = (RMSNorm if config.fuse_norm else nn.RMSNorm)(config.hidden_size, eps=config.norm_eps)
+        if config.attn is not None and layer_idx in config.attn['layers']:
+            self.attn = Attention(
+                hidden_size=config.hidden_size,
+                num_heads=config.attn['num_heads'],
+                num_kv_heads=config.attn['num_kv_heads'],
+                qkv_bias=config.attn['qkv_bias'],
+                window_size=config.attn['window_size'],
+                rope_theta=config.attn['rope_theta'],
+                max_position_embeddings=config.max_position_embeddings,
+                layer_idx=layer_idx
+            )
+        else:
+            self.attn = LinearAttention(
+                mode=config.attn_mode,
+                hidden_size=config.hidden_size,
+                expand_k=config.expand_k,
+                expand_v=config.expand_v,
+                num_heads=config.num_heads,
+                num_kv_heads=config.num_kv_heads,
+                feature_map=config.feature_map,
+                tie_feature_map_qk=config.tie_feature_map_qk,
+                norm_q=config.norm_q,
+                norm_k=config.norm_k,
+                do_feature_map_norm=config.norm_feature_map,
+                elementwise_affine=config.elementwise_affine,
+                norm_eps=config.norm_eps,
+                layer_idx=layer_idx
+            )
+        self.mlp_norm = (RMSNorm if config.fuse_norm else nn.RMSNorm)(config.hidden_size, eps=config.norm_eps)
+        self.mlp = LinearAttentionMLP(
+            hidden_size=config.hidden_size,
+            hidden_ratio=config.hidden_ratio,
+            intermediate_size=config.intermediate_size,
+            hidden_act=config.hidden_act,
+            fuse_swiglu=config.fuse_swiglu
+        )
+
+    def forward(
+        self,
+        hidden_states: torch.Tensor,
+        attention_mask: Optional[torch.Tensor] = None,
+        past_key_values: Optional[Union[Cache, List[torch.FloatTensor]]] = None,
+        use_cache: Optional[bool] = False,
+        output_attentions: Optional[bool] = False,
+        **kwargs,
+    ) -> Tuple[torch.FloatTensor, Optional[Tuple[torch.FloatTensor, torch.FloatTensor]]]:
+        residual = hidden_states
+        # currently not supported
+        attentions, past_key_values = None, None
+        hidden_states = self.attn_norm(hidden_states)
+        hidden_states = self.attn(hidden_states=hidden_states, **kwargs)
+        if self.config.fuse_norm:
+            hidden_states, residual = self.mlp_norm(hidden_states, residual, True)
+        else:
+            hidden_states = residual + hidden_states
+            residual = hidden_states
+            hidden_states = self.mlp_norm(hidden_states)
+        hidden_states = self.mlp(hidden_states, **kwargs)
+        hidden_states = residual + hidden_states
+
+        outputs = (hidden_states, attentions, past_key_values)
+
+        return outputs
+
+
+class LinearAttentionPreTrainedModel(PreTrainedModel):
+
+    config_class = LinearAttentionConfig
+    base_model_prefix = 'model'
+    supports_gradient_checkpointing = True
+    _no_split_modules = ['LinearAttentionBlock']
+    _supports_cache_class = True
+
+    def __init__(self, *inputs, **kwargs):
+        super().__init__(*inputs, **kwargs)
+
+    def _init_weights(
+        self,
+        module: nn.Module,
+        prenorm_residual_strategy: Optional[str] = 'rescale',
+        num_residuals_per_layer: int = 2,
+    ):
+        if isinstance(module, (nn.Linear, nn.Conv1d)):
+            # Slightly different from the TF version which uses truncated_normal for initialization
+            # cf https://github.com/pytorch/pytorch/pull/5617
+            nn.init.normal_(module.weight, mean=0.0, std=self.config.initializer_range)
+            if module.bias is not None:
+                nn.init.zeros_(module.bias)
+        elif isinstance(module, nn.Embedding):
+            nn.init.normal_(module.weight, mean=0.0, std=self.config.initializer_range)
+        elif hasattr(module, 'reset_parameters'):
+            module.reset_parameters()
+
+        if prenorm_residual_strategy is not None:
+            # Reinitialize selected weights subject to the OpenAI GPT-2 Paper Scheme:
+            #   > A modified initialization which accounts for the accumulation on the residual path with model depth. Scale
+            #   > the weights of residual layers at initialization by a factor of 1/√N where N is the # of residual layers.
+            #   >   -- GPT-2 :: https://openai.com/blog/better-language-models/
+            #
+            # Reference (Megatron-LM): https://github.com/NVIDIA/Megatron-LM/blob/main/megatron/model/gpt_model.py
+            p = None
+            if hasattr(module, 'o_proj'):
+                p = module.o_proj.weight
+            elif hasattr(module, 'down_proj'):
+                p = module.down_proj.weight
+            if p is not None:
+                # Special Scaled Initialization --> There are 2 Layer Norms per Transformer Block
+                # Following Pytorch init, except scale by 1/sqrt(2 * n_layer)
+                # We need to reinit p since this code could be called multiple times
+                # Having just p *= scale would repeatedly scale it down
+                if prenorm_residual_strategy == 'rescale':
+                    nn.init.kaiming_uniform_(p, a=math.sqrt(5))
+                    with torch.no_grad():
+                        p /= math.sqrt(num_residuals_per_layer * self.config.num_hidden_layers)
+                elif prenorm_residual_strategy == 'zero':
+                    nn.init.zeros_(p)
+                else:
+                    raise ValueError(f"Invalid prenorm_residual_strategy: {prenorm_residual_strategy}")
+
+
+class LinearAttentionModel(LinearAttentionPreTrainedModel):
+
+    def __init__(self, config: LinearAttentionConfig):
+        super().__init__(config)
+        self.padding_idx = config.pad_token_id
+        self.vocab_size = config.vocab_size
+
+        self.embeddings = nn.Embedding(config.vocab_size, config.hidden_size, self.padding_idx)
+        self.layers = nn.ModuleList([LinearAttentionBlock(config, layer_idx) for layer_idx in range(config.num_hidden_layers)])
+        self.norm = (RMSNorm if config.fuse_norm else nn.RMSNorm)(config.hidden_size, eps=config.norm_eps)
+
+        self.gradient_checkpointing = False
+
+        self.post_init()
+
+    def get_input_embeddings(self):
+        return self.embeddings
+
+    def set_input_embeddings(self, value):
+        self.embeddings = value
+
+    def forward(
+        self,
+        input_ids: Optional[torch.LongTensor] = None,
+        attention_mask: Optional[torch.Tensor] = None,  # noqa
+        inputs_embeds: Optional[torch.FloatTensor] = None,
+        past_key_values: Optional[Union[Cache, List[torch.FloatTensor]]] = None,
+        use_cache: Optional[bool] = None,
+        output_attentions: Optional[bool] = None,
+        output_hidden_states: Optional[bool] = None,
+        return_dict: Optional[bool] = None
+    ) -> Union[Tuple, BaseModelOutputWithPast]:
+        if output_attentions:
+            warnings.warn(
+                "`LinearAttentionModel` does not support output attention weights now, "
+                "so `output_attentions` is set to `False`."
+            )
+            output_attentions = False
+        output_attentions = output_attentions if output_attentions is not None else self.config.output_attentions
+        output_hidden_states = output_hidden_states if output_hidden_states is not None else self.config.output_hidden_states
+        use_cache = use_cache if use_cache is not None else (self.config.use_cache if not self.training else False)
+        return_dict = return_dict if return_dict is not None else self.config.use_return_dict
+
+        # retrieve input_ids and inputs_embeds
+        if input_ids is not None and inputs_embeds is not None:
+            raise ValueError("You cannot specify both input_ids and inputs_embeds at the same time")
+        if input_ids is None and inputs_embeds is None:
+            raise ValueError("You have to specify either input_ids or inputs_embeds")
+
+        if inputs_embeds is None:
+            inputs_embeds = self.embeddings(input_ids)
+        hidden_states = inputs_embeds
+
+        if use_cache and not isinstance(past_key_values, Cache):
+            past_key_values = Cache.from_legacy_cache(past_key_values)
+
+        if self.gradient_checkpointing and self.training and use_cache:
+            logger.warning_once("`use_cache=True` is incompatible with gradient checkpointing. Setting `use_cache=False`...")
+            use_cache = False
+
+        all_hidden_states = () if output_hidden_states else None
+        all_attns = () if output_attentions else None
+
+        for i, layer in enumerate(self.layers):
+            if output_hidden_states:
+                all_hidden_states += (hidden_states,)
+
+            if self.gradient_checkpointing and self.training:
+                hidden_states, attentions, past_key_values = self._gradient_checkpointing_func(
+                    layer.__call__,
+                    hidden_states,
+                    attention_mask,
+                    past_key_values,
+                    use_cache,
+                    output_attentions,
+                )
+            else:
+                hidden_states, attentions, past_key_values = layer(
+                    hidden_states,
+                    attention_mask=attention_mask,
+                    past_key_values=past_key_values,
+                    use_cache=use_cache,
+                    output_attentions=output_attentions
+                )
+
+            if output_attentions:
+                all_attns += (attentions,)
+
+        hidden_states = self.norm(hidden_states)
+
+        # add hidden states from the last decoder layer
+        if output_hidden_states:
+            all_hidden_states += (hidden_states,)
+
+        if not return_dict:
+            return tuple(i for i in [hidden_states, past_key_values, all_hidden_states, all_attns] if i is not None)
+        return BaseModelOutputWithPast(
+            last_hidden_state=hidden_states,
+            past_key_values=past_key_values,
+            hidden_states=all_hidden_states,
+            attentions=all_attns
+        )
+
+
+class LinearAttentionForCausalLM(LinearAttentionPreTrainedModel, GenerationMixin):
+
+    _tied_weights_keys = ["lm_head.weight"]
+
+    def __init__(self, config):
+        super().__init__(config)
+        self.model = LinearAttentionModel(config)
+        self.vocab_size = config.vocab_size
+        self.lm_head = nn.Linear(config.hidden_size, config.vocab_size, bias=False)
+        self.criterion = None
+
+        # Initialize weights and apply final processing
+        self.post_init()
+
+    def get_input_embeddings(self):
+        return self.model.embeddings
+
+    def set_input_embeddings(self, value):
+        self.model.embeddings = value
+
+    def get_output_embeddings(self):
+        return self.lm_head
+
+    def set_output_embeddings(self, new_embeddings):
+        self.lm_head = new_embeddings
+
+    def set_decoder(self, decoder):
+        self.model = decoder
+
+    def get_decoder(self):
+        return self.model
+
+    def generate(self, *args, **kwargs):
+        try:
+            return super().generate(*args, **kwargs)
+        except AttributeError as exception:
+            if 'past_key_values' in str(exception):
+                raise AttributeError(
+                    f"You tried to call `generate` with a decoding strategy that manipulates `past_key_values`, "
+                    f"which is not supported for {self.__class__.__name__}. "
+                    f"Try another generation strategy instead. "
+                    f"For the available generation strategies, check this doc: "
+                    f"https://huggingface.co/docs/transformers/en/generation_strategies#decoding-strategies"
+                )
+            else:
+                raise exception
+
+    @deprecate_kwarg("num_logits_to_keep", version="4.50", new_name="logits_to_keep")
+    def prepare_inputs_for_generation(
+        self,
+        input_ids: torch.LongTensor = None,
+        past_key_values: Optional[Union[Cache, List[torch.FloatTensor]]] = None,
+        attention_mask: Optional[torch.Tensor] = None,
+        inputs_embeds: Optional[torch.Tensor] = None,
+        use_cache: bool = True,
+        logits_to_keep: Optional[int] = None,
+        **kwargs
+    ):
+        # only last token for `inputs_ids` if the `past_key_values` is not empty.
+        if past_key_values is not None and len(past_key_values) > 0:
+            input_ids = input_ids[:, -1:]
+        # if `inputs_embeds` are passed, we only want to use them in the 1st generation step
+        if inputs_embeds is not None and len(past_key_values) == 0:
+            model_inputs = {'inputs_embeds': inputs_embeds}
+        else:
+            # The `contiguous()` here is necessary to have a static stride during decoding. torchdynamo otherwise
+            # recompiles graphs as the stride of the inputs is a guard.
+            # Ref: https://github.com/huggingface/transformers/pull/29114
+            # TODO: use `next_tokens` directly instead.
+            model_inputs = {'input_ids': input_ids.contiguous()}
+
+        if logits_to_keep is not None:
+            model_inputs['logits_to_keep'] = logits_to_keep
+
+        model_inputs.update({
+            'past_key_values': past_key_values,
+            'use_cache': use_cache,
+            'attention_mask': attention_mask,
+        })
+        return model_inputs
+
+    @deprecate_kwarg("num_logits_to_keep", version="4.50", new_name="logits_to_keep")
+    def forward(
+        self,
+        input_ids: torch.LongTensor = None,
+        attention_mask: Optional[torch.Tensor] = None,
+        inputs_embeds: Optional[torch.Tensor] = None,
+        past_key_values: Optional[Union[Cache, List[torch.FloatTensor]]] = None,
+        labels: Optional[torch.LongTensor] = None,
+        use_cache: Optional[bool] = None,
+        output_attentions: Optional[bool] = None,
+        output_hidden_states: Optional[bool] = None,
+        return_dict: Optional[bool] = None,
+        logits_to_keep: Optional[int] = 0
+    ) -> Union[Tuple, CausalLMOutputWithPast]:
+        output_attentions = output_attentions if output_attentions is not None else self.config.output_attentions
+        output_hidden_states = (
+            output_hidden_states if output_hidden_states is not None else self.config.output_hidden_states
+        )
+        return_dict = return_dict if return_dict is not None else self.config.use_return_dict
+
+        outputs = self.model(
+            input_ids=input_ids,
+            attention_mask=attention_mask,
+            inputs_embeds=inputs_embeds,
+            past_key_values=past_key_values,
+            use_cache=use_cache,
+            output_attentions=output_attentions,
+            output_hidden_states=output_hidden_states,
+            return_dict=return_dict
+        )
+
+        hidden_states = outputs[0]
+        fuse_linear_and_cross_entropy = self.config.fuse_cross_entropy and self.training
+
+        loss, logits = None, None
+        if not fuse_linear_and_cross_entropy or labels is None:
+            logits = self.lm_head(hidden_states if logits_to_keep is None else hidden_states[:, -logits_to_keep:])
+        if labels is not None:
+            if getattr(self, 'criterion', None) is None:
+                if fuse_linear_and_cross_entropy:
+                    criterion = FusedLinearCrossEntropyLoss()
+                elif self.config.fuse_cross_entropy:
+                    criterion = FusedCrossEntropyLoss(inplace_backward=True)
+                else:
+                    criterion = nn.CrossEntropyLoss()
+            else:
+                criterion = self.criterion
+            labels = labels.to(hidden_states.device)
+            labels = torch.cat((labels[..., 1:], torch.full_like(labels[:, :1], criterion.ignore_index)), 1)
+            if fuse_linear_and_cross_entropy:
+                loss = criterion(hidden_states, labels, self.lm_head.weight, self.lm_head.bias)
+            else:
+                loss = criterion(logits.view(labels.numel(), -1), labels.view(-1))
+
+        if not return_dict:
+            output = (logits,) + outputs[1:]
+            return (loss,) + output if loss is not None else output
+
+        return CausalLMOutputWithPast(
+            loss=loss,
+            logits=logits,
+            past_key_values=outputs.past_key_values,
+            hidden_states=outputs.hidden_states,
+            attentions=outputs.attentions,
+        )

fla/models/mamba/__init__.py

@@ -0,0 +1,13 @@
+# -*- coding: utf-8 -*-
+
+from transformers import AutoConfig, AutoModel, AutoModelForCausalLM
+
+from fla.models.mamba.configuration_mamba import MambaConfig
+from fla.models.mamba.modeling_mamba import MambaBlock, MambaForCausalLM, MambaModel
+
+AutoConfig.register(MambaConfig.model_type, MambaConfig, True)
+AutoModel.register(MambaConfig, MambaModel, True)
+AutoModelForCausalLM.register(MambaConfig, MambaForCausalLM, True)
+
+
+__all__ = ['MambaConfig', 'MambaForCausalLM', 'MambaModel', 'MambaBlock']

fla/models/mamba/modeling_mamba.py

@@ -0,0 +1,843 @@
+# coding=utf-8
+# Copyright 2024 state-spaces/mamba org and HuggingFace Inc. team.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+"""PyTorch MAMBA model."""
+
+import math
+import warnings
+from dataclasses import dataclass
+from typing import Any, Dict, Optional, Tuple, Union
+
+import torch
+import torch.utils.checkpoint
+from torch import nn
+from transformers.activations import ACT2FN
+from transformers.configuration_utils import PretrainedConfig
+from transformers.generation import GenerationMixin
+from transformers.modeling_utils import PreTrainedModel
+from transformers.utils import ModelOutput, logging
+from transformers.utils.deprecation import deprecate_kwarg
+
+from fla.models.mamba.configuration_mamba import MambaConfig
+from fla.modules import FusedCrossEntropyLoss, FusedLinearCrossEntropyLoss, RMSNorm
+
+logger = logging.get_logger(__name__)
+
+
+with warnings.catch_warnings():
+    warnings.simplefilter('ignore')
+    try:
+        from mamba_ssm.ops.selective_scan_interface import mamba_inner_fn, selective_scan_fn
+        from mamba_ssm.ops.triton.selective_state_update import selective_state_update
+    except ImportError:
+        selective_state_update, selective_scan_fn, mamba_inner_fn = None, None, None
+
+    try:
+        from causal_conv1d import causal_conv1d_fn, causal_conv1d_update
+    except ImportError:
+        causal_conv1d_update, causal_conv1d_fn = None, None
+    is_fast_path_available = all((
+        selective_state_update,
+        selective_scan_fn,
+        causal_conv1d_fn,
+        causal_conv1d_update,
+        mamba_inner_fn
+    ))
+
+
+class MambaCache:
+    """
+    Cache for mamba model which does not have attention mechanism and key value states.
+
+    Arguments:
+        config (`PretrainedConfig):
+            The configuration file defining the shape-related attributes required to initialize the static cache.
+        batch_size (`int`):
+            The batch size with which the model will be used. Note that a new instance must be instantiated if a
+            smaller batch size is used.
+        dtype (`torch.dtype`, *optional*, defaults to `torch.float16`):
+            The default `dtype` to use when initializing the layer.
+        device (`torch.device` or `str`, *optional*):
+            The device on which the cache should be initialized. Should be the same as the layer.
+
+    Attributes:
+        dtype: (`torch.dtype`):
+            The default `dtype` used to initializing the cache.
+        intermediate_size: (`int`):
+            Model's intermediate_size taken from config.
+        ssm_state_size: (`int`):
+            Model's state_size taken from config.
+        conv_kernel_size: (`int`):
+            Model's convolution kernel size taken from config
+        conv_states: (`torch.Tensor`):
+            A tensor of shape `[layer_idx, batch_size, intermediate_size, conv_kernel_size]` that holds convolutional states.
+        ssm_states: (`torch.Tensor`):
+            A tensor of shape `[layer_idx, batch_size, intermediate_size, ssm_state_size]` that holds ssm states
+
+    Example:
+
+        ```python
+        >>> from transformers import AutoTokenizer, MambaForCausalLM, MambaCache
+
+        >>> model = MambaForCausalLM.from_pretrained("state-spaces/mamba-130m-hf")
+        >>> tokenizer = AutoTokenizer.from_pretrained("state-spaces/mamba-130m-hf")
+
+        >>> inputs = tokenizer(text="My name is Mamba", return_tensors="pt")
+
+        >>> # Prepare a cache class and pass it to model's forward
+        >>> past_key_values = MambaCache(config=model.config, batch_size=1, device=model.device, dtype=model.dtype)
+        >>> outputs = model(**inputs, past_key_values=past_key_values, use_cache=True)
+        >>> outputs.past_key_values
+        MambaCache()
+        ```
+    """
+
+    # TODO (joao): remove `=None` in non-optional arguments in v4.46. Remove from `OBJECTS_TO_IGNORE` as well.
+    def __init__(
+        self,
+        config: PretrainedConfig,
+        batch_size: int = None,
+        dtype: torch.dtype = torch.float16,
+        device: Optional[Union[torch.device, str]] = None,
+        max_batch_size: Optional[int] = None,
+    ):
+        if max_batch_size is not None:
+            logger.warning_once(
+                f"The 'max_batch_size' argument of {self.__class__.__name__} is deprecated and will be removed in "
+                "v4.46. Use the more precisely named 'batch_size' argument instead."
+            )
+        self.dtype = dtype
+        self.batch_size = batch_size or max_batch_size
+        self.intermediate_size = config.intermediate_size
+        self.ssm_state_size = config.state_size
+        self.conv_kernel_size = config.conv_kernel
+
+        self.conv_states: torch.Tensor = torch.zeros(
+            config.num_hidden_layers,
+            self.batch_size,
+            self.intermediate_size,
+            self.conv_kernel_size,
+            device=device,
+            dtype=dtype,
+        )
+        self.ssm_states: torch.Tensor = torch.zeros(
+            config.num_hidden_layers,
+            self.batch_size,
+            self.intermediate_size,
+            self.ssm_state_size,
+            device=device,
+            dtype=dtype,
+        )
+
+        torch._dynamo.mark_static_address(self.conv_states)
+        torch._dynamo.mark_static_address(self.ssm_states)
+
+    def update_conv_state(
+        self, layer_idx: int, new_conv_state: torch.Tensor, cache_position: torch.LongTensor
+    ) -> torch.Tensor:
+        conv_state = self.conv_states[layer_idx]
+        cache_position = cache_position.clamp(0, self.conv_kernel_size - 1)
+
+        conv_state = conv_state.roll(shifts=-1, dims=-1)
+        conv_state[:, :, cache_position] = new_conv_state.to(conv_state.device)
+        self.conv_states[layer_idx].zero_()
+        self.conv_states[layer_idx] += conv_state
+        return self.conv_states[layer_idx]
+
+    def update_ssm_state(self, layer_idx: int, new_ssm_state: torch.Tensor):
+        self.ssm_states[layer_idx] = new_ssm_state.to(self.ssm_states.device)
+        return self.ssm_states[layer_idx]
+
+    def reset(self):
+        self.conv_states.zero_()
+        self.ssm_states.zero_()
+
+
+class MambaMixer(nn.Module):
+    """
+    Compute ∆, A, B, C, and D the state space parameters and compute the `contextualized_states`.
+    A, D are input independent (see Mamba paper [1] Section 3.5.2 "Interpretation of A" for why A isn't selective)
+    ∆, B, C are input-dependent (this is a key difference between Mamba and the linear time invariant S4,
+    and is why Mamba is called **selective** state spaces)
+    """
+
+    def __init__(self, config: MambaConfig, layer_idx: int):
+        super().__init__()
+        self.config = config
+        self.hidden_size = config.hidden_size
+        self.ssm_state_size = config.state_size
+        self.conv_kernel_size = config.conv_kernel
+        self.intermediate_size = config.intermediate_size
+        self.time_step_rank = int(config.time_step_rank)
+        self.layer_idx = layer_idx
+        self.use_conv_bias = config.use_conv_bias
+        self.conv1d = nn.Conv1d(
+            in_channels=self.intermediate_size,
+            out_channels=self.intermediate_size,
+            bias=config.use_conv_bias,
+            kernel_size=config.conv_kernel,
+            groups=self.intermediate_size,
+            padding=config.conv_kernel - 1,
+        )
+
+        self.activation = config.hidden_act
+        self.act = ACT2FN[config.hidden_act]
+
+        # projection of the input hidden states
+        self.in_proj = nn.Linear(self.hidden_size, self.intermediate_size * 2, bias=config.use_bias)
+        # selective projection used to make dt, B and C input dependant
+        self.x_proj = nn.Linear(self.intermediate_size, self.time_step_rank + self.ssm_state_size * 2, bias=False)
+        # time step projection (discretization)
+        self.dt_proj = nn.Linear(self.time_step_rank, self.intermediate_size, bias=True)
+
+        # S4D real initialization. These are not discretized!
+        # The core is to load them, compute the discrete states, then write the updated state. Keeps the memory bounded
+        A = torch.arange(1, self.ssm_state_size + 1, dtype=torch.float32)[None, :]
+        A = A.expand(self.intermediate_size, -1).contiguous()
+
+        self.A_log = nn.Parameter(torch.log(A))
+        self.D = nn.Parameter(torch.ones(self.intermediate_size))
+        self.out_proj = nn.Linear(self.intermediate_size, self.hidden_size, bias=config.use_bias)
+        self.use_bias = config.use_bias
+
+        if not is_fast_path_available:
+            logger.warning_once(
+                "The fast path is not available because on of "
+                "`(selective_state_update, selective_scan_fn, causal_conv1d_fn, causal_conv1d_update, mamba_inner_fn)`"
+                " is None. Falling back to the naive implementation. "
+                "To install follow https://github.com/state-spaces/mamba/#installation and"
+                " https://github.com/Dao-AILab/causal-conv1d"
+            )
+
+    def cuda_kernels_forward(
+        self,
+        hidden_states: torch.Tensor,
+        cache_params: Optional[MambaCache] = None,
+        cache_position: Optional[torch.LongTensor] = None,
+        attention_mask: Optional[torch.LongTensor] = None,
+    ):
+        # 1. Gated MLP's linear projection
+        projected_states = self.in_proj(hidden_states).transpose(1, 2)
+
+        if self.training and cache_params is None:  # Doesn't support outputting the states -> used for training
+            contextualized_states = mamba_inner_fn(
+                projected_states,
+                self.conv1d.weight,
+                self.conv1d.bias if self.use_conv_bias else None,
+                self.x_proj.weight,
+                self.dt_proj.weight,
+                self.out_proj.weight,
+                self.out_proj.bias.float() if self.use_bias else None,
+                -torch.exp(self.A_log.float()),
+                None,  # input-dependent B
+                None,  # input-dependent C
+                self.D.float(),
+                delta_bias=self.dt_proj.bias.float(),
+                delta_softplus=True,
+            )
+
+        else:
+            hidden_states, gate = projected_states.chunk(2, dim=1)
+
+            if attention_mask is not None:
+                hidden_states = hidden_states * attention_mask.unsqueeze(1)
+
+            # 2. Convolution sequence transformation
+            conv_weights = self.conv1d.weight.view(self.conv1d.weight.size(0), self.conv1d.weight.size(2))
+            if cache_params is not None and cache_position[0] > 0:
+                hidden_states = causal_conv1d_update(
+                    hidden_states.squeeze(-1),
+                    cache_params.conv_states[self.layer_idx],
+                    conv_weights,
+                    self.conv1d.bias,
+                    self.activation,
+                )
+                hidden_states = hidden_states.unsqueeze(-1)
+            else:
+                if cache_params is not None:
+                    conv_states = nn.functional.pad(
+                        hidden_states, (self.conv_kernel_size - hidden_states.shape[-1], 0)
+                    )
+                    cache_params.update_conv_state(self.layer_idx, conv_states, cache_position)
+                hidden_states = causal_conv1d_fn(
+                    hidden_states, conv_weights, self.conv1d.bias, activation=self.activation
+                )
+
+            if attention_mask is not None:
+                hidden_states = hidden_states * attention_mask.unsqueeze(1)
+
+            # 3. State Space Model sequence transformation
+            # 3.a. input varying initialization of time_step, B and C
+            ssm_parameters = self.x_proj(hidden_states.transpose(1, 2))
+            time_step, B, C = torch.split(
+                ssm_parameters, [self.time_step_rank, self.ssm_state_size, self.ssm_state_size], dim=-1
+            )
+            discrete_time_step = self.dt_proj.weight @ time_step.transpose(1, 2)
+
+            A = -torch.exp(self.A_log.float())
+            # 3.c perform the recurrence y ← SSM(A, B, C)(x)
+            time_proj_bias = self.dt_proj.bias.float() if hasattr(self.dt_proj, "bias") else None
+            if cache_params is not None and cache_position[0] > 0:
+                scan_outputs = selective_state_update(
+                    cache_params.ssm_states[self.layer_idx],
+                    hidden_states[..., 0],
+                    discrete_time_step[..., 0],
+                    A,
+                    B[:, 0],
+                    C[:, 0],
+                    self.D,
+                    gate[..., 0],
+                    time_proj_bias,
+                    dt_softplus=True,
+                ).unsqueeze(-1)
+            else:
+                scan_outputs, ssm_state = selective_scan_fn(
+                    hidden_states,
+                    discrete_time_step,
+                    A,
+                    B.transpose(1, 2),
+                    C.transpose(1, 2),
+                    self.D.float(),
+                    gate,
+                    time_proj_bias,
+                    delta_softplus=True,
+                    return_last_state=True,
+                )
+                if ssm_state is not None and cache_params is not None:
+                    cache_params.update_ssm_state(self.layer_idx, ssm_state)
+
+            # 4. Final linear projection
+            contextualized_states = self.out_proj(scan_outputs.transpose(1, 2))
+        return contextualized_states
+
+    def slow_forward(
+        self,
+        input_states,
+        cache_params: Optional[MambaCache] = None,
+        cache_position: Optional[torch.LongTensor] = None,
+        attention_mask: Optional[torch.LongTensor] = None
+    ):
+        batch_size, seq_len, _ = input_states.shape
+        dtype = input_states.dtype
+        # 1. Gated MLP's linear projection
+        # [batch, 2 * intermediate_size, seq_len]
+        projected_states = self.in_proj(input_states).transpose(1, 2)
+        hidden_states, gate = projected_states.chunk(2, dim=1)
+
+        if attention_mask is not None:
+            hidden_states = hidden_states * attention_mask.unsqueeze(1)
+
+        # 2. Convolution sequence transformation
+        if cache_params is not None:
+            ssm_state = cache_params.ssm_states[self.layer_idx].clone()
+            ssm_state = ssm_state.to(hidden_states.device)
+            # use `cache_position.shape[0]` to check whether we are in prefill
+            # stage, it's equivalent to check `cache_position[0] == 0`, which
+            # breaks dynamo fullgraph constraints
+            if cache_position.shape[0] == self.conv_kernel_size:
+                conv_state = nn.functional.pad(
+                    hidden_states,
+                    (self.conv_kernel_size - hidden_states.shape[-1], 0)
+                )
+
+                cache_params.update_conv_state(self.layer_idx, conv_state, cache_position)
+                # [batch, intermediate_size, seq_len]
+                hidden_states = self.act(self.conv1d(hidden_states)[..., :seq_len])
+            else:
+                conv_state = cache_params.update_conv_state(self.layer_idx, hidden_states, cache_position)
+                hidden_states = torch.sum(conv_state * self.conv1d.weight[:, 0, :], dim=-1)
+                if self.use_conv_bias:
+                    hidden_states += self.conv1d.bias
+                # [batch, intermediate_size, 1] : decoding
+                hidden_states = self.act(hidden_states).to(dtype).unsqueeze(-1)
+        else:
+            ssm_state = torch.zeros(
+                (batch_size, self.intermediate_size, self.ssm_state_size),
+                device=hidden_states.device, dtype=dtype
+            )
+            # [batch, intermediate_size, seq_len]
+            hidden_states = self.act(self.conv1d(hidden_states)[..., :seq_len])
+
+        if attention_mask is not None:
+            hidden_states = hidden_states * attention_mask.unsqueeze(1)
+
+        # 3. State Space Model sequence transformation
+        # 3.a. Selection:  [batch, seq_len, self.time_step_rank + self.ssm_state_size * 2]
+        ssm_parameters = self.x_proj(hidden_states.transpose(1, 2))
+        time_step, B, C = torch.split(
+            ssm_parameters, [self.time_step_rank, self.ssm_state_size, self.ssm_state_size], dim=-1
+        )
+        # [batch, seq_len, intermediate_size]
+        discrete_time_step = self.dt_proj(time_step)
+        # [batch, intermediate_size, seq_len]
+        discrete_time_step = nn.functional.softplus(discrete_time_step).transpose(1, 2)
+
+        # 3.b. Discretization: B and C to [batch, seq_len, intermediate_size, ssm_state_size] (SRAM)
+        # [intermediate_size, ssm_state_size]
+        A = -torch.exp(self.A_log.float())
+        # [batch, intermediate_size, seq_len, ssm_state_size]
+        discrete_A = torch.exp(A[None, :, None, :] * discrete_time_step[:, :, :, None])
+        # [batch, intermediate_size, seq_len, ssm_state_size]
+        discrete_B = discrete_time_step[:, :, :, None] * B[:, None, :, :].float()
+        deltaB_u = discrete_B * hidden_states[:, :, :, None].float()
+
+        # 3.c perform the recurrence y ← SSM(A, B, C)(x)
+        scan_outputs = []
+        for i in range(seq_len):
+            # [batch, intermediade_size, ssm_state]
+            ssm_state = discrete_A[:, :, i, :] * ssm_state + deltaB_u[:, :, i, :]
+            # [batch, intermediade_size, 1]
+            scan_output = torch.matmul(ssm_state.to(dtype), C[:, i, :].unsqueeze(-1))
+            scan_outputs.append(scan_output[:, :, 0])
+        # [batch, seq_len, intermediade_size]
+        scan_output = torch.stack(scan_outputs, dim=-1)
+        scan_output = scan_output + (hidden_states * self.D[None, :, None])
+        scan_output = (scan_output * self.act(gate))
+
+        if cache_params is not None:
+            cache_params.ssm_states[self.layer_idx].copy_(ssm_state)
+
+        # 4. Final linear projection
+        # [batch, seq_len, hidden_size]
+        contextualized_states = self.out_proj(scan_output.transpose(1, 2))
+        return contextualized_states
+    # fmt: on
+
+    def forward(
+        self,
+        hidden_states,
+        cache_params: Optional[MambaCache] = None,
+        cache_position: Optional[torch.LongTensor] = None,
+        attention_mask: Optional[torch.LongTensor] = None,
+    ):
+        if is_fast_path_available and "cuda" in self.x_proj.weight.device.type:
+            return self.cuda_kernels_forward(hidden_states, cache_params, cache_position, attention_mask)
+        return self.slow_forward(hidden_states, cache_params, cache_position, attention_mask)
+
+
+class MambaBlock(nn.Module):
+    def __init__(self, config, layer_idx):
+        super().__init__()
+        self.config = config
+        self.layer_idx = layer_idx
+        self.residual_in_fp32 = config.residual_in_fp32
+        self.norm = RMSNorm(config.hidden_size, eps=config.layer_norm_epsilon)
+        self.mixer = MambaMixer(config, layer_idx=layer_idx)
+
+    def forward(
+        self,
+        hidden_states,
+        cache_params: Optional[MambaCache] = None,
+        cache_position: Optional[torch.LongTensor] = None,
+        attention_mask: Optional[torch.LongTensor] = None,
+    ):
+        residual = hidden_states
+        hidden_states = self.norm(hidden_states)
+        if self.residual_in_fp32:
+            residual = residual.to(torch.float32)
+
+        hidden_states = self.mixer(
+            hidden_states, cache_params=cache_params, cache_position=cache_position, attention_mask=attention_mask
+        )
+        hidden_states = residual + hidden_states
+        if self.residual_in_fp32:
+            hidden_states = hidden_states.to(dtype=self.norm.weight.dtype)
+        return hidden_states
+
+
+class MambaPreTrainedModel(PreTrainedModel):
+    """
+    An abstract class to handle weights initialization and a simple interface for downloading and loading pretrained
+    models.
+    """
+
+    config_class = MambaConfig
+    base_model_prefix = "backbone"
+    _no_split_modules = ["MambaBlock", "MambaMixer"]
+    supports_gradient_checkpointing = True
+    _is_stateful = True
+
+    def _init_weights(self, module):
+        """Initialize the weights."""
+        if isinstance(module, nn.Linear):
+            nn.init.normal_(module.weight, mean=0.0, std=self.config.initializer_range)
+            if module.bias is not None:
+                if not getattr(module.bias, "_no_reinit", False):
+                    nn.init.zeros_(module.bias)
+        elif isinstance(module, MambaMixer):
+            module.A_log._no_weight_decay = True
+            module.D._no_weight_decay = True
+
+            dt_init_std = self.config.time_step_rank**-0.5 * self.config.time_step_scale
+            if self.config.time_step_init_scheme == "constant":
+                nn.init.constant_(module.dt_proj.weight, dt_init_std)
+            elif self.config.time_step_init_scheme == "random":
+                nn.init.uniform_(module.dt_proj.weight, -dt_init_std, dt_init_std)
+
+            dt = torch.exp(
+                torch.rand(self.config.intermediate_size)
+                * (math.log(self.config.time_step_max) - math.log(self.config.time_step_min))
+                + math.log(self.config.time_step_min)
+            ).clamp(min=self.config.time_step_floor)
+            # # Inverse of softplus: https://github.com/pytorch/pytorch/issues/72759
+            inv_dt = dt + torch.log(-torch.expm1(-dt))
+            with torch.no_grad():
+                module.dt_proj.bias.data = nn.Parameter(inv_dt.to(module.dt_proj.bias.device))
+            module.dt_proj.bias._no_reinit = True
+        elif isinstance(module, nn.Embedding):
+            nn.init.normal_(module.weight, std=self.config.initializer_range)
+        elif hasattr(module, 'reset_parameters'):
+            module.reset_parameters()
+
+        if self.config.rescale_prenorm_residual:
+            # Reinitialize selected weights subject to the OpenAI GPT-2 Paper Scheme:
+            #   > A modified initialization which accounts for the accumulation on the residual path with model depth. Scale
+            #   > the weights of residual layers at initialization by a factor of 1/√N where N is the # of residual layers.
+            #   >   -- GPT-2 :: https://openai.com/blog/better-language-models/
+            #
+            # Reference (Megatron-LM): https://github.com/NVIDIA/Megatron-LM/blob/main/megatron/model/gpt_model.py
+            for name, p in module.named_parameters():
+                if name in ["out_proj.weight"]:
+                    # Special Scaled Initialization --> There are 2 Layer Norms per Transformer Block
+                    # Following Pytorch init, except scale by 1/sqrt(2 * n_layer)
+                    # We need to reinit p since this code could be called multiple times
+                    # Having just p *= scale would repeatedly scale it down
+                    nn.init.kaiming_uniform_(p, a=math.sqrt(5))
+                    with torch.no_grad():
+                        p /= math.sqrt(self.config.num_hidden_layers)
+
+
+@dataclass
+class MambaOutput(ModelOutput):
+    """
+    Class for the MAMBA model outputs.
+
+    Args:
+        last_hidden_state (`torch.FloatTensor` of shape `(batch_size, sequence_length, hidden_size)`):
+            Sequence of hidden-states at the output of the last layer of the model.
+        cache_params (`MambaCache`):
+            The state of the model at the last time step. Can be used in a forward method with the next `input_ids` to
+            avoid providing the old `input_ids`.
+
+            Includes both the State space model state matrices after the selective scan, and the Convolutional states
+        hidden_states (`tuple(torch.FloatTensor)`, *optional*,
+            returned when `output_hidden_states=True` is passed or when `config.output_hidden_states=True`):
+            Tuple of `torch.FloatTensor` (one for the output of the embeddings, if the model has an embedding layer, +
+            one for the output of each layer) of shape `(batch_size, sequence_length, hidden_size)`.
+
+            Hidden-states of the model at the output of each layer plus the optional initial embedding outputs.
+    """
+
+    last_hidden_state: Optional[torch.FloatTensor] = None
+    cache_params: Optional[MambaCache] = None
+    hidden_states: Optional[Tuple[torch.FloatTensor]] = None
+
+
+@dataclass
+class MambaCausalLMOutput(ModelOutput):
+    """
+    Base class for causal language model (or autoregressive) outputs.
+
+    Args:
+        loss (`torch.FloatTensor` of shape `(1,)`, *optional*, returned when `labels` is provided):
+            Language modeling loss (for next-token prediction).
+        logits (`torch.FloatTensor` of shape `(batch_size, sequence_length, config.vocab_size)`):
+            Prediction scores of the language modeling head (scores for each vocabulary token before SoftMax).
+        cache_params (`MambaCache`):
+            The state of the model at the last time step. Can be used in a forward method with the next `input_ids` to
+            avoid providing the old `input_ids`.
+
+            Includes both the State space model state matrices after the selective scan, and the Convolutional states
+        hidden_states (`tuple(torch.FloatTensor)`, *optional*,
+            returned when `output_hidden_states=True` is passed or when `config.output_hidden_states=True`):
+            Tuple of `torch.FloatTensor` (one for the output of the embeddings, if the model has an embedding layer, +
+            one for the output of each layer) of shape `(batch_size, sequence_length, hidden_size)`.
+
+            Hidden-states of the model at the output of each layer plus the optional initial embedding outputs.
+    """
+
+    loss: Optional[torch.FloatTensor] = None
+    logits: Optional[torch.FloatTensor] = None
+    cache_params: Optional[MambaCache] = None
+    hidden_states: Optional[Tuple[torch.FloatTensor]] = None
+
+
+class MambaModel(MambaPreTrainedModel):
+    def __init__(self, config):
+        super().__init__(config)
+
+        self.embeddings = nn.Embedding(config.vocab_size, config.hidden_size)
+        self.layers = nn.ModuleList([MambaBlock(config, layer_idx=idx) for idx in range(config.num_hidden_layers)])
+
+        self.gradient_checkpointing = False
+        self.norm_f = RMSNorm(config.hidden_size, eps=config.layer_norm_epsilon)
+        # Initialize weights and apply final processing
+        self._register_load_state_dict_pre_hook(self.load_hook)
+        self.post_init()
+
+    def load_hook(self, state_dict, prefix, *args):
+        for k in state_dict:
+            if "embedding." in k:
+                state_dict[k.replace("embedding.", "embeddings.")] = state_dict.pop(k)
+                break
+
+    def get_input_embeddings(self):
+        return self.embeddings
+
+    def set_input_embeddings(self, new_embeddings):
+        self.embeddings = new_embeddings
+
+    def forward(
+        self,
+        input_ids: Optional[torch.LongTensor] = None,
+        inputs_embeds: Optional[torch.LongTensor] = None,
+        cache_params: Optional[MambaCache] = None,
+        use_cache: Optional[bool] = None,
+        output_hidden_states: Optional[bool] = None,
+        return_dict: Optional[bool] = None,
+        cache_position: Optional[torch.LongTensor] = None,
+        attention_mask: Optional[torch.LongTensor] = None,
+    ) -> Union[Tuple, MambaOutput]:
+        output_hidden_states = (
+            output_hidden_states if output_hidden_states is not None else self.config.output_hidden_states
+        )
+        use_cache = use_cache if use_cache is not None else (self.config.use_cache if not self.training else False)
+        return_dict = return_dict if return_dict is not None else self.config.use_return_dict
+
+        if (input_ids is None) ^ (inputs_embeds is not None):  # ^ is python for xor
+            raise ValueError(
+                "You cannot specify both input_ids and inputs_embeds at the same time, and must specify either one"
+            )
+
+        if inputs_embeds is None:
+            inputs_embeds = self.embeddings(input_ids)
+
+        if self.gradient_checkpointing and self.training and use_cache:
+            use_cache = False
+
+        if use_cache:
+            if cache_params is None:
+                cache_params = MambaCache(
+                    self.config, inputs_embeds.size(0), device=inputs_embeds.device, dtype=inputs_embeds.dtype
+                )
+                cache_position = torch.arange(0, self.config.conv_kernel, device=inputs_embeds.device)
+            elif cache_position is None:
+                # cases when we do manual forward instead of using `model.generate` which will initiate
+                # `cache_position` and makes sure it is not None, throw error here instead of doing some
+                # hack to conjecture the current cache position
+                raise ValueError(
+                    "You have to specify the `cache_position` manually when `use_cache=True` and `cache_params` is passed, "
+                    "you don't have to pass a `cache_params` if you are in prefilling stage because in that case it will "
+                    "be initialized for you automatically"
+                )
+        else:
+            cache_params = None
+
+        hidden_states = inputs_embeds
+        all_hidden_states = () if output_hidden_states else None
+        for mixer_block in self.layers:
+            if self.gradient_checkpointing and self.training:
+                hidden_states = self._gradient_checkpointing_func(
+                    mixer_block.__call__, hidden_states, cache_params, cache_position, attention_mask
+                )
+            else:
+                hidden_states = mixer_block(
+                    hidden_states,
+                    cache_params=cache_params,
+                    cache_position=cache_position,
+                    attention_mask=attention_mask,
+                )
+
+            if output_hidden_states:
+                all_hidden_states = all_hidden_states + (hidden_states,)
+
+        hidden_states = self.norm_f(hidden_states)
+
+        if output_hidden_states:
+            all_hidden_states = all_hidden_states + (hidden_states,)
+
+        if not return_dict:
+            return tuple(v for v in [hidden_states, cache_params, all_hidden_states] if v is not None)
+
+        return MambaOutput(
+            last_hidden_state=hidden_states,
+            cache_params=cache_params if use_cache else None,
+            hidden_states=all_hidden_states,
+        )
+
+
+class MambaForCausalLM(MambaPreTrainedModel, GenerationMixin):
+
+    _tied_weights_keys = ["lm_head.weight"]
+
+    def __init__(self, config):
+        super().__init__(config)
+        self.backbone = MambaModel(config)
+        self.lm_head = nn.Linear(config.hidden_size, config.vocab_size, bias=False)
+        self.criterion = None
+
+        # Initialize weights and apply final processing
+        self.post_init()
+
+    def get_output_embeddings(self):
+        return self.lm_head
+
+    def set_output_embeddings(self, new_embeddings):
+        self.lm_head = new_embeddings
+
+    def get_input_embeddings(self):
+        return self.backbone.get_input_embeddings()
+
+    def set_input_embeddings(self, new_embeddings):
+        return self.backbone.set_input_embeddings(new_embeddings)
+
+    def _update_model_kwargs_for_generation(
+        self, outputs: ModelOutput,
+        model_kwargs: Dict[str, Any],
+        num_new_tokens: int = 1,
+        **kwargs
+    ) -> Dict[str, Any]:
+        model_kwargs["cache_params"] = outputs.get("cache_params", None)
+        if (
+            model_kwargs.get("use_cache", True)
+            and "cache_position" in model_kwargs
+            and model_kwargs["cache_position"] is not None
+        ):
+            model_kwargs["cache_position"] = model_kwargs["cache_position"][-1:] + num_new_tokens
+
+        if "attention_mask" in model_kwargs:
+            attention_mask = model_kwargs["attention_mask"]
+            model_kwargs["attention_mask"] = torch.cat(
+                [attention_mask, attention_mask.new_ones((attention_mask.shape[0], 1))], dim=-1
+            )
+
+        return model_kwargs
+
+    @deprecate_kwarg("num_logits_to_keep", version="4.50", new_name="logits_to_keep")
+    def prepare_inputs_for_generation(
+        self,
+        input_ids,
+        inputs_embeds=None,
+        use_cache=None,
+        cache_params: Optional[MambaCache] = None,
+        cache_position: Optional[torch.LongTensor] = None,
+        attention_mask: Optional[torch.LongTensor] = None,
+        logits_to_keep: Optional[int] = None,
+        **kwargs,
+    ):
+        if use_cache:
+            # `cache_position` should have been initialized in `generate`
+            if cache_position is None:
+                raise ValueError(
+                    "`cache_position` should not be None as it should have been initialized in "
+                    "`model.generate`, you are responsible for passing in a valid `cache_position` if "
+                    "you are calling `prepare_inputs_for_generation` directly with `use_cache=True`"
+                )
+            if cache_position[0] > 0:
+                input_ids = input_ids[:, -1].unsqueeze(-1)
+
+                if attention_mask is not None:
+                    attention_mask = None
+
+            else:
+                # we initialize the `cache_position` to full size of `conv_states` at prefill stage
+                # considering padding will be applied when input length is shorter, and truncation
+                # will be applied when it is longer, so it will be equivalent to always have it match
+                # the length of `cache_params.conv_states`, which is `config.conv_kernel`
+                cache_position = torch.arange(0, self.config.conv_kernel, device=input_ids.device)
+
+        if inputs_embeds is not None and cache_params is None:
+            model_inputs = {"inputs_embeds": inputs_embeds}
+        else:
+            model_inputs = {"input_ids": input_ids.contiguous()}
+
+        if logits_to_keep is not None:
+            model_inputs['logits_to_keep'] = logits_to_keep
+
+        model_inputs.update({
+            'cache_params': cache_params,
+            'use_cache': use_cache,
+            'cache_position': cache_position,
+            'attention_mask': attention_mask,
+            'logits_to_keep': logits_to_keep,
+        })
+        return model_inputs
+
+    @deprecate_kwarg("num_logits_to_keep", version="4.50", new_name="logits_to_keep")
+    def forward(
+        self,
+        input_ids: Optional[torch.LongTensor] = None,
+        attention_mask: Optional[torch.LongTensor] = None,
+        inputs_embeds: Optional[torch.FloatTensor] = None,
+        cache_params: Optional[MambaCache] = None,
+        labels: Optional[torch.LongTensor] = None,
+        output_hidden_states: Optional[bool] = None,
+        return_dict: Optional[bool] = None,
+        use_cache: Optional[bool] = None,
+        cache_position: Optional[torch.Tensor] = None,
+        logits_to_keep: Optional[int] = 0,
+        **kwargs,  # for now we need this for generation
+    ) -> Union[Tuple, MambaCausalLMOutput]:
+        r"""
+        labels (`torch.LongTensor` of shape `(batch_size, sequence_length)`, *optional*):
+            Labels for language modeling. Note that the labels **are shifted** inside the model, i.e. you can set
+            `labels = input_ids` Indices are selected in `[-100, 0, ..., config.vocab_size]` All labels set to `-100`
+            are ignored (masked), the loss is only computed for labels in `[0, ..., config.vocab_size]`
+        """
+        return_dict = return_dict if return_dict is not None else self.config.use_return_dict
+
+        mamba_outputs = self.backbone(
+            input_ids,
+            cache_params=cache_params,
+            inputs_embeds=inputs_embeds,
+            output_hidden_states=output_hidden_states,
+            return_dict=return_dict,
+            use_cache=use_cache,
+            cache_position=cache_position,
+            attention_mask=attention_mask,
+        )
+        hidden_states = mamba_outputs[0]
+        fuse_linear_and_cross_entropy = self.config.fuse_cross_entropy and self.training
+
+        loss, logits = None, None
+        if not fuse_linear_and_cross_entropy or labels is None:
+            logits = self.lm_head(hidden_states if logits_to_keep is None else hidden_states[:, -logits_to_keep:])
+        if labels is not None:
+            if getattr(self, 'criterion', None) is None:
+                if fuse_linear_and_cross_entropy:
+                    criterion = FusedLinearCrossEntropyLoss()
+                elif self.config.fuse_cross_entropy:
+                    criterion = FusedCrossEntropyLoss(inplace_backward=True)
+                else:
+                    criterion = nn.CrossEntropyLoss()
+            else:
+                criterion = self.criterion
+            # Enable model parallelism
+            labels = labels.to(hidden_states.device)
+            labels = torch.cat((labels[..., 1:], torch.full_like(labels[:, :1], criterion.ignore_index)), 1)
+            if fuse_linear_and_cross_entropy:
+                loss = criterion(hidden_states, labels, self.lm_head.weight, self.lm_head.bias)
+            else:
+                loss = criterion(logits.view(labels.numel(), -1), labels.view(-1))
+
+        if not return_dict:
+            output = (logits,) + mamba_outputs[1:]
+            return (loss,) + output if loss is not None else output
+
+        return MambaCausalLMOutput(
+            loss=loss,
+            logits=logits,
+            cache_params=mamba_outputs.cache_params,
+            hidden_states=mamba_outputs.hidden_states,
+        )

fla/models/mamba2/configuration_mamba2.py

@@ -0,0 +1,170 @@
+# Copyright 2024 The HuggingFace Inc. team.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+"""MAMBA2 configuration"""
+
+import math
+
+from transformers.configuration_utils import PretrainedConfig
+
+
+class Mamba2Config(PretrainedConfig):
+    """
+    This is the configuration class to store the configuration of a [`Mamba2Model`]. It is used to instantiate a MAMBA2
+    model according to the specified arguments, defining the model architecture. Instantiating a configuration with the
+    defaults will yield a similar configuration to that of the MAMBA2
+    [state-spaces/mamba2-2.8b](https://huggingface.co/state-spaces/mamba2-2.8b) architecture.
+
+    Configuration objects inherit from [`PretrainedConfig`] and can be used to control the model outputs. Read the
+    documentation from [`PretrainedConfig`] for more information.
+
+
+    Args:
+        num_heads (`int`, *optional*, defaults to 64):
+            Number of heads for the evolution matrices of mamba 2.
+        head_dim (`int`, *optional*, defaults to 64):
+            Dimension of each head.
+        vocab_size (`int`, *optional*, defaults to 32768):
+            Vocabulary size of the MAMBA2 model. Defines the number of different tokens that can be represented by the
+            `inputs_ids` passed when calling [`Mamba2Model`].
+        hidden_size (`int`, *optional*, defaults to 2048):
+            Dimensionality of the embeddings and hidden states.
+        state_size (`int`, *optional*, defaults to 128): shape of the state space latents.
+        num_hidden_layers (`int`, *optional*, defaults to 48):
+            Number of hidden layers in the model.
+        layer_norm_epsilon (`float`, *optional*, defaults to 1e-05):
+            The epsilon to use in the layer normalization layers.
+        pad_token_id (`int`, *optional*, defaults to 0):
+            Padding token id.
+        bos_token_id (`int`, *optional*, defaults to 1):
+            The id of the beginning of sentence token in the vocabulary.
+        eos_token_id (`int`, *optional*, defaults to 2):
+            The id of the end of sentence token in the vocabulary.
+        expand (`int`, *optional*, defaults to 2): Expanding factor used to determine the intermediate size.
+        conv_kernel (`int`, *optional*, defaults to 4): Size of the convolution kernel.
+        n_groups (`int`, *optional*, defaults to 1):
+            Number of groups for the evolution matrices of mamba 2.
+        use_bias (`bool`, *optional*, defaults to `False`):
+            Whether or not to use bias in ["in_proj", "out_proj"] of the mixer block
+        use_conv_bias (`bool`, *optional*, defaults to `True`):
+            Whether or not to use bias in the convolution layer of the mixer block.
+        hidden_act (`str`, *optional*, defaults to `"silu"`):
+            The non-linear activation function (function or string) in the decoder.
+        initializer_range (`float`, *optional*, defaults to 0.1):
+            The standard deviation of the truncated_normal_initializer for initializing all weight matrices.
+        residual_in_fp32 (`bool`, *optional*, defaults to `True`):
+            Whether or not residuals should be in `float32`.
+            If set to `False` residuals will keep the same `dtype` as the rest of the model
+        time_step_rank (`Union[int,str]`, *optional*, defaults to `"auto"`):
+            Rank of the discretization projection matrix.
+            `"auto"` means that it will default to `math.ceil(self.hidden_size / 16)`
+        time_step_min (`float`, *optional*, defaults to 0.001):
+            Minimum `time_step` used to bound `dt_proj.bias`.
+        time_step_max (`float`, *optional*, defaults to 0.1):
+            Maximum `time_step` used to bound `dt_proj.bias`.
+        time_step_floor (`float`, *optional*, defaults to 0.0001):
+            Minimum clamping value of the `dt_proj.bias` layer initialization.
+        time_step_limit (`tuple`, *optional*, defaults to `(0.0, inf)`):
+            Accepted range of time step values.
+        rescale_prenorm_residual (`bool`, *optional*, defaults to `True`):
+            Whether or not to rescale `out_proj` weights when initializing.
+        use_cache (`bool`, *optional*, defaults to `True`):
+            Whether or not the cache should be used.
+        rms_norm (`bool`, *optional*, defaults to `True`):
+            Whether to use RMS norm or not.
+        chunk_size (`int`, *optional*, defaults to 256):
+            Size of the chunks that will comprise the sequence.
+        tie_word_embeddings (`bool`, *optional*, defaults to `False`):
+            Whether to tie word embeddings or not.
+    """
+
+    model_type = "mamba2"
+
+    def __init__(
+        self,
+        num_heads: int = 64,
+        head_dim: int = 64,
+        vocab_size: int = 32000,
+        hidden_size: int = 2048,
+        state_size: int = 128,
+        num_hidden_layers: int = 48,
+        layer_norm_epsilon: float = 1e-5,
+        pad_token_id: int = 0,
+        bos_token_id: int = 1,
+        eos_token_id: int = 2,
+        expand: int = 2,
+        conv_kernel: int = 4,
+        n_groups: int = 1,
+        use_bias: bool = False,
+        use_conv_bias: bool = True,
+        hidden_act: str = "silu",
+        initializer_range: float = 0.1,
+        residual_in_fp32: bool = True,
+        time_step_rank: str = "auto",
+        time_step_min: float = 0.001,
+        time_step_max: float = 0.1,
+        time_step_floor: float = 1e-4,
+        time_step_limit=(0.0, float("inf")),
+        rescale_prenorm_residual: bool = True,
+        use_cache: bool = True,
+        rms_norm: bool = True,
+        chunk_size: int = 256,
+        fuse_norm: bool = True,
+        fuse_cross_entropy: bool = True,
+        tie_word_embeddings: bool = False,
+        **kwargs,
+    ):
+        self.vocab_size = vocab_size
+        self.hidden_size = hidden_size
+        self.state_size = state_size
+        self.num_hidden_layers = num_hidden_layers
+        self.layer_norm_epsilon = layer_norm_epsilon
+        self.conv_kernel = conv_kernel
+        self.expand = expand
+
+        self.bos_token_id = bos_token_id
+        self.eos_token_id = eos_token_id
+        self.pad_token_id = pad_token_id
+        self.use_bias = use_bias
+        self.use_conv_bias = use_conv_bias
+        self.hidden_act = hidden_act
+        self.initializer_range = initializer_range
+        self.time_step_rank = (
+            math.ceil(self.hidden_size / 16)
+            if time_step_rank == "auto"
+            else time_step_rank
+        )
+        self.time_step_min = time_step_min
+        self.time_step_max = time_step_max
+        self.time_step_floor = time_step_floor
+        self.rescale_prenorm_residual = rescale_prenorm_residual
+        self.residual_in_fp32 = residual_in_fp32
+        self.use_cache = use_cache
+        self.n_groups = n_groups
+        self.num_heads = num_heads
+        self.head_dim = head_dim
+        self.rms_norm = rms_norm
+        self.state_size = state_size
+        self.chunk_size = chunk_size
+        self.time_step_limit = time_step_limit
+        self.fuse_norm = fuse_norm
+        self.fuse_cross_entropy = fuse_cross_entropy
+        self.tie_word_embeddings = tie_word_embeddings
+
+        super().__init__(
+            bos_token_id=bos_token_id,
+            eos_token_id=eos_token_id,
+            pad_token_id=pad_token_id,
+            tie_word_embeddings=tie_word_embeddings,
+            **kwargs,
+        )

fla/models/mamba2/modeling_mamba2.py

@@ -0,0 +1,1093 @@
+# Copyright 2024 state-spaces/mamba2 org and HuggingFace Inc. team.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+"""PyTorch MAMBA2 model."""
+
+import math
+import warnings
+from dataclasses import dataclass
+from typing import Optional, Tuple, Union
+
+import torch
+import torch.utils.checkpoint
+from torch import nn
+from transformers.activations import ACT2FN
+from transformers.generation import GenerationMixin
+from transformers.modeling_utils import PreTrainedModel
+from transformers.utils import ModelOutput, logging
+from transformers.utils.deprecation import deprecate_kwarg
+
+from fla.models.mamba2.configuration_mamba2 import Mamba2Config
+from fla.modules import FusedCrossEntropyLoss, FusedLinearCrossEntropyLoss, RMSNorm
+from fla.modules.layernorm_gated import RMSNormGated
+
+logger = logging.get_logger(__name__)
+
+with warnings.catch_warnings():
+    warnings.simplefilter('ignore')
+    try:
+        from mamba_ssm.ops.triton.selective_state_update import selective_state_update
+        from mamba_ssm.ops.triton.ssd_combined import mamba_chunk_scan_combined, mamba_split_conv1d_scan_combined
+    except ImportError:
+        (
+            selective_state_update,
+            mamba_chunk_scan_combined,
+            mamba_split_conv1d_scan_combined,
+        ) = (None, None, None)
+    try:
+        from causal_conv1d import causal_conv1d_fn, causal_conv1d_update
+    except ImportError:
+        causal_conv1d_update, causal_conv1d_fn = None, None
+    is_fast_path_available = all((
+        selective_state_update,
+        causal_conv1d_fn,
+        causal_conv1d_update
+    ))
+
+
+def pad_tensor_by_size(input_tensor: torch.Tensor, pad_size: int):
+    """
+    Padding x tensor with `pad_size` on the seq_len dim (dim=1)
+
+    Assumes that we only have tensors of either size 4 or 3
+    """
+    pad_shape = (0, 0, 0, 0, 0, pad_size, 0, 0) if len(input_tensor.shape) == 4 else (0, 0, 0, pad_size, 0, 0)
+
+    return torch.nn.functional.pad(input_tensor, pad_shape, mode="constant", value=0)
+
+
+def reshape_into_chunks(input_tensor, pad_size, chunk_size):
+    """
+    Padding input_tensor with `pad_size` on the seq_len dim (dim=1) and
+    simultaneously splitting it into chunk sequences.
+
+    Assumes that we only have tensors of either size 4 or 3
+    """
+    # [bsz, seq_len, ...] -> [bsz, seq_len multiple of chunk_size, ...]
+    input_tensor = pad_tensor_by_size(input_tensor, pad_size)
+
+    if len(input_tensor.shape) == 3:
+        # [bsz, seq_len multiple of chunk_size, num_heads] -> [bsz, -1, chunk_size, num_heads]
+        return input_tensor.reshape(input_tensor.shape[0], -1, chunk_size, input_tensor.shape[2])
+    else:
+        # [bsz, seq_len multiple of chunk_size, num_heads, head_dim or state_size] ->
+        # [bsz, -1, chunk_size, num_heads, head_dim or state_size]
+        return input_tensor.reshape(
+            input_tensor.shape[0], -1, chunk_size, input_tensor.shape[2], input_tensor.shape[3]
+        )
+
+
+def segment_sum(input_tensor):
+    """
+    More stable segment sum calculation. Uses cumulative sums and masking instead of direct subtractions.
+    """
+    chunk_size = input_tensor.size(-1)
+    # 1. expand input tensor to have an additional dimension and repeat along that dimension
+    # [..., chunk_size] -> [..., chunk_size, chunk_size]
+    input_tensor = input_tensor[..., None].expand(*input_tensor.size(), chunk_size)
+    # 2. create a lower triangular mask with the diagonal set to 0 to 0 out elements above diag
+    mask = torch.tril(torch.ones(chunk_size, chunk_size, device=input_tensor.device, dtype=torch.bool), diagonal=-1)
+    input_tensor = input_tensor.masked_fill(~mask, 0)
+    # 3. compute actual cumsum
+    tensor_segsum = torch.cumsum(input_tensor, dim=-2)
+
+    # 4. apply mask to keep only the lower triangular part of the cumulative sum result (incl diagonal this time)
+    mask = torch.tril(torch.ones(chunk_size, chunk_size, device=input_tensor.device, dtype=torch.bool), diagonal=0)
+    tensor_segsum = tensor_segsum.masked_fill(~mask, -torch.inf)
+    return tensor_segsum
+
+
+def apply_mask_to_padding_states(hidden_states, attention_mask):
+    """
+    Tunes out the hidden states for padding tokens, see https://github.com/state-spaces/mamba/issues/66
+    """
+    if attention_mask is not None and attention_mask.shape[1] > 1 and attention_mask.shape[0] > 1:
+        dtype = hidden_states.dtype
+        hidden_states = (hidden_states * attention_mask[:, :, None]).to(dtype)
+
+    return hidden_states
+
+
+class Mamba2Cache:
+    """
+    Arguments:
+        config: Mamba2Config
+        batch_size: int
+        dtype: torch.dtype
+        device: torch.device
+
+    Attributes:
+        dtype: (`torch.dtype`):
+            The default `dtype` used to initializing the cache.
+        conv_kernel_size: (`int`):
+            Model's convolution kernel size taken from config.
+        n_groups: (`int`):
+            Model's number of groups taken from the config - similar to tensor parallel in Transformer.
+        state_size: (`int`):
+            Model's SSM state size taken from config.
+        num_heads: (`int`):
+            The number of heads used in the linear attention / SSM.
+        head_dim: (`int`):
+            The respective dimension of the heads used in the linear attention / SSM.
+        intermediate_size: (`int`):
+            Model's intermediate_size based on (expand * hidden_dim) from config.
+        conv_states: (`torch.Tensor`):
+            A tensor of shape `[num_layers, batch_size, conv_kernel_size, intermediate_size + 2 * n_groups * state_size]`
+            that holds convolutional states.
+        ssm_states: (`torch.Tensor`):
+            A tensor of shape `[num_layers, batch_size, num_heads, head_dim, state_size]` that holds ssm states.
+    """
+
+    def __init__(
+        self,
+        config: Mamba2Config,
+        batch_size: int,
+        dtype: torch.dtype = torch.float16,
+        device: Optional[str] = None,
+    ):
+        self.dtype = dtype
+        self.conv_kernel_size = config.conv_kernel
+        self.n_groups = config.n_groups
+        self.state_size = config.state_size
+        self.num_heads = config.num_heads
+        self.head_dim = config.head_dim
+        self.intermediate_size = int(config.expand * config.hidden_size)
+
+        self.conv_states = torch.zeros(
+            config.num_hidden_layers,
+            batch_size,
+            self.intermediate_size + 2 * self.n_groups * self.state_size,
+            self.conv_kernel_size,
+            device=device,
+            dtype=dtype,
+        )
+        self.ssm_states = torch.zeros(
+            config.num_hidden_layers,
+            batch_size,
+            self.num_heads,
+            self.head_dim,
+            self.state_size,
+            device=device,
+            dtype=dtype,
+        )
+
+    def update_conv_state(
+        self,
+        layer_idx: int,
+        new_conv_state: torch.Tensor,
+        cache_init: bool = False
+    ) -> torch.Tensor:
+        if cache_init:
+            self.conv_states[layer_idx] = new_conv_state.to(self.conv_states.device)
+        else:
+            self.conv_states[layer_idx] = self.conv_states[layer_idx].roll(shifts=-1, dims=-1)
+            self.conv_states[layer_idx][:, :, -1] = new_conv_state[:, 0, :].to(self.conv_states.device)
+        return self.conv_states[layer_idx]
+
+    def update_ssm_state(self, layer_idx: int, new_ssm_state: torch.Tensor):
+        self.ssm_states[layer_idx] = new_ssm_state.to(self.ssm_states.device)
+        return self.ssm_states[layer_idx]
+
+    def reset(self):
+        self.conv_states.zero_()
+        self.ssm_states.zero_()
+
+
+class Mamba2Mixer(nn.Module):
+    """
+    Compute ∆, A, B, C, and D the state space parameters and compute the `contextualized_states`.
+    A, D are input independent (see Mamba paper [1] Section 3.5.2 "Interpretation of A" for why A isn't selective)
+    ∆, B, C are input-dependent (this is a key difference between Mamba and the linear time invariant S4,
+    and is why Mamba is called **selective** state spaces)
+    """
+
+    def __init__(self, config: Mamba2Config, layer_idx: int):
+        super().__init__()
+        self.num_heads = config.num_heads
+        self.hidden_size = config.hidden_size
+        self.ssm_state_size = config.state_size
+        self.conv_kernel_size = config.conv_kernel
+        self.intermediate_size = int(config.expand * self.hidden_size)
+        self.time_step_rank = int(config.time_step_rank)
+        self.layer_idx = layer_idx
+        self.use_conv_bias = config.use_conv_bias
+        self.activation = config.hidden_act
+        self.act = ACT2FN[config.hidden_act]
+
+        self.layer_norm_epsilon = config.layer_norm_epsilon
+        self.rms_norm = config.rms_norm
+
+        self.n_groups = config.n_groups
+        self.head_dim = config.head_dim
+        self.chunk_size = config.chunk_size
+
+        self.time_step_limit = config.time_step_limit
+        self.time_step_min = config.time_step_min
+        self.time_step_max = config.time_step_max
+
+        self.conv_dim = self.intermediate_size + 2 * self.n_groups * self.ssm_state_size
+        self.conv1d = nn.Conv1d(
+            in_channels=self.conv_dim,
+            out_channels=self.conv_dim,
+            bias=config.use_conv_bias,
+            kernel_size=config.conv_kernel,
+            groups=self.conv_dim,
+            padding=config.conv_kernel - 1,
+        )
+
+        # projection of the input hidden states
+        projection_size = self.intermediate_size + self.conv_dim + self.num_heads
+        self.in_proj = nn.Linear(
+            self.hidden_size,
+            projection_size,
+            bias=config.use_bias,
+        )
+        # selective projection used to make dt, B and C input dependant
+
+        # time step projection (discretization)
+        # instantiate once and copy inv_dt in init_weights of PretrainedModel
+        self.dt_bias = nn.Parameter(torch.ones(self.num_heads))
+
+        # S4D real initialization. These are not discretized!
+        # The core is to load them, compute the discrete states, then write the updated state. Keeps the memory bounded
+        A = torch.arange(1, self.num_heads + 1)
+        self.A_log = nn.Parameter(torch.log(A))
+        self.A_log._no_weight_decay = True
+        self.norm = RMSNormGated(
+            self.intermediate_size, eps=self.layer_norm_epsilon, norm_before_gate=False
+        )
+        self.D = nn.Parameter(torch.ones(self.num_heads))
+        self.D._no_weight_decay = True
+
+        self.out_proj = nn.Linear(self.intermediate_size, self.hidden_size, bias=config.use_bias)
+        self.use_bias = config.use_bias
+
+        if not is_fast_path_available:
+            logger.warning_once(
+                "The fast path is not available because one of "
+                "`(selective_state_update, causal_conv1d_fn, causal_conv1d_update)` is None. "
+                "Falling back to the naive implementation. "
+                "To install follow https://github.com/state-spaces/mamba/#installation and"
+                "https://github.com/Dao-AILab/causal-conv1d"
+            )
+
+    def cuda_kernels_forward(
+        self,
+        hidden_states: torch.Tensor,
+        cache_params: Optional[Mamba2Cache] = None,
+        cache_position: Optional[torch.LongTensor] = None,
+        attention_mask: Optional[torch.Tensor] = None,
+    ):
+        # 1. Gated MLP's linear projection
+        hidden_states = apply_mask_to_padding_states(hidden_states, attention_mask)
+        projected_states = self.in_proj(hidden_states)
+
+        # Set up dimensions for reshapes later
+        batch_size, seq_len, _ = hidden_states.shape
+        groups_time_state_size = self.n_groups * self.ssm_state_size
+        d_mlp = (
+            projected_states.shape[-1]
+            - 2 * self.intermediate_size
+            - 2 * self.n_groups * self.ssm_state_size
+            - self.num_heads
+        ) // 2
+
+        # Single step calculations via cache
+        if cache_params is not None and cache_position is not None and cache_position[0] > 0:
+            _, _, gate, hidden_states_B_C, dt = projected_states.squeeze(1).split(
+                [d_mlp, d_mlp, self.intermediate_size, self.conv_dim, self.num_heads], dim=-1
+            )
+
+            # 2. Convolution sequence transformation
+            hidden_states_B_C = causal_conv1d_update(
+                hidden_states_B_C,
+                cache_params.conv_states[self.layer_idx],
+                self.conv1d.weight.squeeze(1),
+                self.conv1d.bias,
+                self.activation,
+            )
+
+            hidden_states, B, C = torch.split(
+                hidden_states_B_C,
+                [
+                    self.intermediate_size,
+                    groups_time_state_size,
+                    groups_time_state_size,
+                ],
+                dim=-1,
+            )
+
+            # 3. SSM transformation
+            A = -torch.exp(self.A_log.float())  # (nheads,)
+            A = A[:, None, ...][:, :, None].expand(-1, self.head_dim, self.ssm_state_size).to(dtype=torch.float32)
+            dt = dt[:, :, None].expand(-1, -1, self.head_dim)
+            dt_bias = self.dt_bias[:, None, ...].expand(-1, self.head_dim)
+            D = self.D[:, None, ...].expand(-1, self.head_dim)
+            B = B.view(batch_size, self.n_groups, B.shape[1] // self.n_groups)
+            C = C.view(batch_size, self.n_groups, C.shape[1] // self.n_groups)
+            hidden_states_reshaped = hidden_states.view(batch_size, self.num_heads, self.head_dim)
+
+            hidden_states = selective_state_update(
+                cache_params.ssm_states[self.layer_idx],
+                hidden_states_reshaped,
+                dt,
+                A,
+                B,
+                C,
+                D,
+                z=None,
+                dt_bias=dt_bias,
+                dt_softplus=True,
+            )
+            hidden_states = hidden_states.view(batch_size, self.num_heads * self.head_dim)
+            hidden_states = self.norm(hidden_states, gate)
+
+            # 4. Final linear projection
+            out = self.out_proj(hidden_states)[:, None, ...]
+
+        # Fused calculations or step by step if no initialized cache is found
+        else:
+            A = -torch.exp(self.A_log.float())  # (num_heads) or (intermediate_size, state_size)
+            dt_limit_kwargs = {} if self.time_step_limit == (0.0, float("inf")) else {"dt_limit": self.time_step_limit}
+
+            # 2-4. Fused kernel for conv1d, SSM, and the final projection
+            if self.training and cache_params is None:
+                out = mamba_split_conv1d_scan_combined(
+                    projected_states,
+                    self.conv1d.weight.squeeze(1),
+                    self.conv1d.bias,
+                    self.dt_bias,
+                    A,
+                    D=self.D,
+                    chunk_size=self.chunk_size,
+                    seq_idx=None,  # was seq_idx
+                    activation=self.activation,
+                    rmsnorm_weight=self.norm.weight,
+                    rmsnorm_eps=self.norm.eps,
+                    outproj_weight=self.out_proj.weight,
+                    outproj_bias=self.out_proj.bias,
+                    headdim=self.head_dim,
+                    ngroups=self.n_groups,
+                    norm_before_gate=False,
+                    return_final_states=False,
+                    **dt_limit_kwargs,
+                )
+
+            else:
+                _, _, gate, hidden_states_B_C, dt = projected_states.split(
+                    [d_mlp, d_mlp, self.intermediate_size, self.conv_dim, self.num_heads], dim=-1
+                )
+
+                # 2. Convolution sequence transformation
+                # Init cache
+                if cache_params is not None:
+                    hidden_states_B_C_transposed = hidden_states_B_C.transpose(1, 2)
+                    conv_states = nn.functional.pad(
+                        hidden_states_B_C_transposed,
+                        (cache_params.conv_kernel_size - hidden_states_B_C_transposed.shape[-1], 0),
+                    )
+                    cache_params.update_conv_state(
+                        layer_idx=self.layer_idx, new_conv_state=conv_states, cache_init=True
+                    )
+
+                if self.activation not in ["silu", "swish"]:
+                    hidden_states_B_C = self.act(
+                        self.conv1d(hidden_states_B_C.transpose(1, 2))[..., :seq_len].transpose(1, 2)
+                    )
+                else:
+                    hidden_states_B_C = causal_conv1d_fn(
+                        x=hidden_states_B_C.transpose(1, 2),
+                        weight=self.conv1d.weight.squeeze(1),
+                        bias=self.conv1d.bias,
+                        activation=self.activation,
+                    ).transpose(1, 2)
+
+                hidden_states_B_C = apply_mask_to_padding_states(hidden_states_B_C, attention_mask)
+                hidden_states, B, C = torch.split(
+                    hidden_states_B_C,
+                    [self.intermediate_size, groups_time_state_size, groups_time_state_size],
+                    dim=-1,
+                )
+
+                # 3. SSM transformation
+                scan_output, ssm_state = mamba_chunk_scan_combined(
+                    hidden_states.view(batch_size, seq_len, -1, self.head_dim),
+                    dt,
+                    A,
+                    B.view(batch_size, seq_len, self.n_groups, -1),
+                    C.view(batch_size, seq_len, self.n_groups, -1),
+                    chunk_size=self.chunk_size,
+                    D=self.D,
+                    z=None,
+                    seq_idx=None,
+                    return_final_states=True,
+                    dt_bias=self.dt_bias,
+                    dt_softplus=True,
+                    **dt_limit_kwargs,
+                )
+
+                # Init cache
+                if ssm_state is not None and cache_params is not None:
+                    cache_params.update_ssm_state(layer_idx=self.layer_idx, new_ssm_state=ssm_state)
+
+                scan_output = scan_output.view(batch_size, seq_len, -1)
+                # Multiply "gate" branch and apply extra normalization layer
+                scan_output = self.norm(scan_output, gate)
+
+                # 4. Final linear projection
+                out = self.out_proj(scan_output)
+        return out
+
+    # fmt: off
+    def torch_forward(
+        self,
+        input_states,
+        cache_params: Optional[Mamba2Cache] = None,
+        cache_position: Optional[torch.LongTensor] = None,
+        attention_mask: Optional[torch.Tensor] = None
+    ):
+        batch_size, seq_len, _ = input_states.shape
+        dtype = input_states.dtype
+
+        # 1. Gated MLP's linear projection
+        input_states = apply_mask_to_padding_states(input_states, attention_mask)
+        projected_states = self.in_proj(input_states)
+        d_mlp = (projected_states.shape[-1] - 2 * self.intermediate_size -
+                 2 * self.n_groups * self.ssm_state_size - self.num_heads) // 2
+        _, _, gate, hidden_states_B_C, dt = projected_states.split(
+            [d_mlp, d_mlp, self.intermediate_size,  self.conv_dim, self.num_heads], dim=-1
+        )
+
+        # 2. Convolution sequence transformation
+        if cache_params is not None and cache_position is not None and cache_position[0] > 0:
+            cache_params.update_conv_state(layer_idx=self.layer_idx, new_conv_state=hidden_states_B_C, cache_init=False)
+
+            # We need to guarantee that anything regarding the cache is on the same device
+            conv_states = cache_params.conv_states[self.layer_idx].to(device=self.conv1d.weight.device)
+
+            hidden_states_B_C = torch.sum(
+                conv_states * self.conv1d.weight.squeeze(1), dim=-1
+            )
+            if self.use_conv_bias:
+                hidden_states_B_C = hidden_states_B_C + self.conv1d.bias
+            hidden_states_B_C = self.act(hidden_states_B_C)
+        else:
+            # Init cache
+            if cache_params is not None:
+                hidden_states_B_C_transposed = hidden_states_B_C.transpose(1, 2)
+                conv_states = nn.functional.pad(
+                    hidden_states_B_C_transposed, (cache_params.conv_kernel_size - hidden_states_B_C_transposed.shape[-1], 0)
+                )
+                cache_params.update_conv_state(layer_idx=self.layer_idx, new_conv_state=conv_states, cache_init=True)
+
+            hidden_states_B_C = self.act(self.conv1d(hidden_states_B_C.transpose(1, 2))[..., :seq_len].transpose(1, 2))
+
+        hidden_states_B_C = apply_mask_to_padding_states(hidden_states_B_C, attention_mask)
+        hidden_states, B, C = torch.split(
+            hidden_states_B_C,
+            [self.intermediate_size, self.n_groups * self.ssm_state_size, self.n_groups * self.ssm_state_size],
+            dim=-1
+        )
+
+        # 3. SSM transformation
+        A = -torch.exp(self.A_log.float())                            # [num_heads]
+        if cache_params is not None and cache_position is not None and cache_position[0] > 0:
+            # We need to guarantee that anything regarding the cache is on the same device
+            cache_device = cache_params.ssm_states.device
+
+            # Note: there is no need to pad parameter matrices here, as there is just one new token
+            # for batched generation
+            dt = dt[:, 0, :][:, None, ...]
+            dt = dt.transpose(1, 2).expand(batch_size, dt.shape[-1], self.head_dim)
+            # [num_heads] -> [num_heads, head_dim]
+            dt_bias = self.dt_bias[..., None].expand(self.dt_bias.shape[0], self.head_dim)
+
+            dt = torch.nn.functional.softplus(dt + dt_bias.to(dt.dtype))
+            dt = torch.clamp(dt, self.time_step_limit[0], self.time_step_limit[1])
+            A = A[..., None, None].expand(self.num_heads, self.head_dim, self.ssm_state_size).to(dtype=torch.float32)
+            # [bsz, num_heads, head_dim, state_size]
+            dA = (torch.exp(dt[..., None] * A)).to(device=cache_device)
+
+            # Discretize B
+            # [bsz, n_groups * state_size] -> [bsz, n_groups, 1, state_size] ->
+            # -> [bsz, n_groups, group to head repetition factor, state_size] -> [bsz, num_heads, state_size]
+            B = B.reshape(batch_size, self.n_groups, -1)[..., None, :]
+            B = B.expand(batch_size, self.n_groups, self.num_heads // self.n_groups, B.shape[-1]).contiguous()
+            B = B.reshape(batch_size, -1, B.shape[-1])
+            # [bsz, num_heads, head_dim, state_size]
+            dB = dt[..., None] * B[..., None, :]
+
+            # Discretize x into dB
+            # [bsz, intermediate_size] -> [bsz, num_heads, head_dim]
+            hidden_states = hidden_states.reshape(batch_size, -1, self.head_dim)
+            dBx = (dB * hidden_states[..., None]).to(device=cache_device)
+
+            # State calculation
+            cache_params.update_ssm_state(
+                layer_idx=self.layer_idx,
+                new_ssm_state=cache_params.ssm_states[self.layer_idx] * dA + dBx
+            )
+
+            # Subsequent output
+            # [bsz, n_groups * state_size] -> [bsz, num_heads, state_size]
+            C = C.reshape(batch_size, self.n_groups, -1)[..., None, :]
+            C = C.expand(batch_size, self.n_groups, self.num_heads // self.n_groups, C.shape[-1]).contiguous()
+            C = C.reshape(batch_size, -1, C.shape[-1])
+            # [bsz, num_heads, head_dim]
+
+            ssm_states = cache_params.ssm_states[self.layer_idx].to(device=C.device, dtype=C.dtype)  # Shape: [b, h, d, n]
+            # Reshape ssm_states to merge the first two dimensions
+            # Shape: [b*h, d, n]
+            ssm_states_reshaped = ssm_states.view(batch_size * self.num_heads, self.head_dim, self.ssm_state_size)
+            C_reshaped = C.view(batch_size * self.num_heads, self.ssm_state_size, 1)  # Shape: [b*h, n, 1]
+            y = torch.bmm(ssm_states_reshaped, C_reshaped)
+            y = y.view(batch_size, self.num_heads, self.head_dim)
+
+            # D skip connection
+            # [num_heads] -> [num_heads, head_dim]
+            D = self.D[..., None].expand(self.D.shape[0], self.head_dim)
+            y = (y + hidden_states * D).to(y.dtype)
+
+            # [bsz, num_heads, head_dim] -> [bsz, 1, intermediate_size]
+            y = y.reshape(batch_size, -1)[:, None, ...]
+        else:
+            # begin ssd naive implementation without einsums
+            dt = nn.functional.softplus(dt + self.dt_bias)
+            dt = torch.clamp(dt, self.time_step_limit[0], self.time_step_limit[1])
+            hidden_states = hidden_states.reshape(batch_size, seq_len, -1, self.head_dim).float()
+            B = B.reshape(batch_size, seq_len, -1, self.ssm_state_size).float()
+            C = C.reshape(batch_size, seq_len, -1, self.ssm_state_size).float()
+            B = B.repeat(1, 1, self.num_heads // self.n_groups, 1)
+            C = C.repeat(1, 1, self.num_heads // self.n_groups, 1)
+            pad_size = (self.chunk_size - seq_len % self.chunk_size) % self.chunk_size
+
+            D_residual = self.D[..., None] * pad_tensor_by_size(hidden_states, pad_size)
+
+            # Discretize x and A
+            hidden_states = hidden_states * dt[..., None]
+            A = A.to(hidden_states.dtype) * dt
+
+            # Rearrange into blocks/chunks
+            hidden_states, A, B, C = [reshape_into_chunks(t, pad_size, self.chunk_size) for t in (hidden_states, A, B, C)]
+
+            # [bsz, -1, chunk_size, num_heads] -> [bsz, num_heads, -1, chunk_size]
+            A = A.permute(0, 3, 1, 2)
+            A_cumsum = torch.cumsum(A, dim=-1)
+
+            # 1. Compute the output for each intra-chunk (diagonal blocks)
+            # This is the analog of a causal mask
+            L = torch.exp(segment_sum(A))
+
+            # Contraction of C and B to get G (attention-weights like)
+            # shape: (b, c, l, s, h, n)
+            G_intermediate = C[:, :, :, None, :, :] * B[:, :, None, :, :, :]
+            G = G_intermediate.sum(dim=-1)  # shape: (b, c, l, s, h)
+
+            # Compute M, equivalent to applying attention mask to weights
+            M_intermediate = G[..., None] * L.permute(0, 2, 3, 4, 1)[..., None]
+            M = M_intermediate.sum(dim=-1)
+
+            # Compute Y_diag (apply to values)
+            Y_diag = (M[..., None] * hidden_states[:, :, None]).sum(dim=3)
+
+            # 2. Compute the state for each intra-chunk
+            # (right term of low-rank factorization of off-diagonal blocks; B terms)
+            decay_states = torch.exp((A_cumsum[:, :, :, -1:] - A_cumsum))
+            B_decay = B * decay_states.permute(0, -2, -1, 1)[..., None]
+            states = (B_decay[..., None, :] * hidden_states[..., None]).sum(dim=2)
+
+            # 3. Compute the inter-chunk SSM recurrence; produces correct SSM states at chunk boundaries
+            # (middle term of factorization of off-diag blocks; A terms)
+            if cache_params is not None and cache_position is not None and cache_position[0] > 0:
+                previous_states = cache_params.ssm_states[self.layer_idx][:, None, ...].to(device=states.device)
+            else:
+                previous_states = torch.zeros_like(states[:, :1])
+            states = torch.cat([previous_states, states], dim=1)
+            decay_chunk = torch.exp(segment_sum(nn.functional.pad(A_cumsum[:, :, :, -1], (1, 0))))
+            decay_chunk = decay_chunk.transpose(1, 3)
+            new_states = (decay_chunk[..., None, None] * states[:, :, None, ...]).sum(dim=1)
+            states, ssm_state = new_states[:, :-1], new_states[:, -1]
+
+            # 4. Compute state -> output conversion per chunk
+            # (left term of low-rank factorization of off-diagonal blocks; C terms)
+            state_decay_out = torch.exp(A_cumsum)
+            C_times_states = (C[..., None, :] * states[:, :, None, ...])
+            state_decay_out_permuted = state_decay_out.permute(0, 2, 3, 1)
+            Y_off = (C_times_states.sum(-1) * state_decay_out_permuted[..., None])
+
+            # Add output of intra-chunk and inter-chunk terms (diagonal and off-diagonal blocks)
+            y = Y_diag + Y_off
+            # [bsz, -1, self.chunk_size, num_heads, head_dim] -> [bsz, (padded) seq_len, num_heads, head_dim]
+            y = y.reshape(batch_size, -1, self.num_heads, self.head_dim)
+
+            y = y + D_residual
+            # Cutting off padded chunks
+            if pad_size > 0:
+                y = y[:, :seq_len, :, :]
+            y = y.reshape(batch_size, seq_len, -1)
+
+            # Init cache
+            if ssm_state is not None and cache_params is not None:
+                cache_params.update_ssm_state(layer_idx=self.layer_idx, new_ssm_state=ssm_state)
+
+        scan_output = self.norm(y, gate)
+
+        # end ssd naive
+
+        # 4. Final linear projection
+        contextualized_states = self.out_proj(scan_output.to(dtype))  # [batch, seq_len, hidden_size]
+        return contextualized_states
+    # fmt: on
+
+    def forward(
+        self,
+        hidden_states,
+        cache_params: Optional[Mamba2Cache] = None,
+        cache_position: Optional[torch.LongTensor] = None,
+        attention_mask: Optional[torch.Tensor] = None,
+    ):
+        if is_fast_path_available and "cuda" in self.in_proj.weight.device.type:
+            return self.cuda_kernels_forward(hidden_states, cache_params, cache_position, attention_mask)
+        dtype = hidden_states.dtype
+        if attention_mask is not None and attention_mask.shape[1] > 1 and attention_mask.shape[0] > 1:
+            # tune out hidden states for pad tokens, see https://github.com/state-spaces/mamba/issues/66
+            hidden_states = (hidden_states * attention_mask[:, :, None]).to(dtype)
+
+        return self.torch_forward(hidden_states, cache_params, cache_position, attention_mask)
+
+
+class Mamba2Block(nn.Module):
+    def __init__(self, config, layer_idx):
+        super().__init__()
+        self.config = config
+        self.layer_idx = layer_idx
+        self.residual_in_fp32 = config.residual_in_fp32
+        self.norm = RMSNorm(config.hidden_size, eps=config.layer_norm_epsilon)
+        self.mixer = Mamba2Mixer(config, layer_idx=layer_idx)
+
+    def forward(
+        self,
+        hidden_states,
+        cache_params: Optional[Mamba2Cache] = None,
+        cache_position: Optional[torch.LongTensor] = None,
+        attention_mask: Optional[torch.Tensor] = None,
+    ):
+        residual = hidden_states
+        hidden_states = self.norm(hidden_states)
+        if self.residual_in_fp32:
+            residual = residual.to(torch.float32)
+
+        hidden_states = self.mixer(
+            hidden_states,
+            cache_params=cache_params,
+            cache_position=cache_position,
+            attention_mask=attention_mask,
+        )
+        hidden_states = residual + hidden_states
+        if self.residual_in_fp32:
+            hidden_states = hidden_states.to(dtype=self.norm.weight.dtype)
+        return hidden_states
+
+
+class Mamba2PreTrainedModel(PreTrainedModel, GenerationMixin):
+    """
+    An abstract class to handle weights initialization and a simple interface for downloading and loading pretrained
+    models.
+    """
+
+    config_class = Mamba2Config
+    base_model_prefix = "backbone"
+    _no_split_modules = ["Mamba2Block"]
+    supports_gradient_checkpointing = True
+    _is_stateful = True
+
+    def _init_weights(
+        self,
+        module: nn.Module,
+        num_residuals_per_layer: int = 1,
+    ):
+        """Initialize the weights."""
+        if isinstance(module, Mamba2Mixer):
+
+            # --- A_log ---
+            A = torch.arange(1, module.num_heads + 1)
+            with torch.no_grad():
+                if not isinstance(module.A_log, torch.distributed.tensor.DTensor):
+                    module.A_log.copy_(torch.log(A))
+                else:
+                    logger.warning_once("`A_log` is a DTensor, skipping initialization")
+            module.A_log._no_weight_decay = True
+
+            # --- D ---
+            nn.init.ones_(module.D)
+            module.D._no_weight_decay = True
+
+            # --- dt_bias ---
+            dt = torch.exp(
+                torch.rand(self.config.num_heads)
+                * (math.log(self.config.time_step_max) - math.log(self.config.time_step_min))
+                + math.log(self.config.time_step_min)
+            ).clamp(min=self.config.time_step_floor)
+
+            # Inverse of softplus: https://github.com/pytorch/pytorch/issues/72759
+            inv_dt = dt + torch.log(-torch.expm1(-dt))
+            with torch.no_grad():
+                if not isinstance(module.dt_bias, torch.distributed.tensor.DTensor):
+                    module.dt_bias.copy_(inv_dt)
+                else:
+                    logger.warning_once("`dt_bias` is a DTensor, skipping initialization")
+            module.dt_bias._no_reinit = True
+
+        elif isinstance(module, (nn.Linear, nn.Conv1d)):
+            # Slightly different from the TF version which uses truncated_normal for initialization
+            # cf https://github.com/pytorch/pytorch/pull/5617
+            nn.init.normal_(module.weight, mean=0.0, std=self.config.initializer_range)
+            if module.bias is not None:
+                nn.init.zeros_(module.bias)
+                # guard against deprecated behavior
+                if hasattr(module.bias, "_no_reinit"):
+                    raise ValueError("This is not supposed to happen")
+        elif isinstance(module, nn.Embedding):
+            nn.init.normal_(module.weight, mean=0.0, std=self.config.initializer_range)
+        elif hasattr(module, 'reset_parameters'):
+            module.reset_parameters()
+
+        if self.config.rescale_prenorm_residual:
+            # Reinitialize selected weights subject to the OpenAI GPT-2 Paper Scheme:
+            #   > A modified initialization which accounts for the accumulation on the residual path with model depth. Scale
+            #   > the weights of residual layers at initialization by a factor of 1/√N where N is the # of residual layers.
+            #   >   -- GPT-2 :: https://openai.com/blog/better-language-models/
+            #
+            # Reference (Megatron-LM): https://github.com/NVIDIA/Megatron-LM/blob/main/megatron/model/gpt_model.py
+            p = None
+            if hasattr(module, 'o_proj'):
+                # p = module.o_proj.weight
+                # guard against deprecated behavior
+                raise ValueError("This is not supposed to happen")
+            elif hasattr(module, 'out_proj'):
+                p = module.out_proj.weight
+            elif hasattr(module, 'down_proj'):
+                p = module.down_proj.weight
+            if p is not None:
+                # Special Scaled Initialization --> There are 2 Layer Norms per Transformer Block
+                # Following Pytorch init, except scale by 1/sqrt(2 * n_layer)
+                # We need to reinit p since this code could be called multiple times
+                # Having just p *= scale would repeatedly scale it down
+                nn.init.kaiming_uniform_(p, a=math.sqrt(5))
+                with torch.no_grad():
+                    p /= math.sqrt(num_residuals_per_layer * self.config.num_hidden_layers)
+
+
+@dataclass
+# Copied from transformers.models.mamba.modeling_mamba.MambaOutput with MAMBA->MAMBA2,Mamba->Mamba2
+class Mamba2Output(ModelOutput):
+    """
+    Class for the MAMBA2 model outputs.
+
+    Args:
+        last_hidden_state (`torch.FloatTensor` of shape `(batch_size, sequence_length, hidden_size)`):
+            Sequence of hidden-states at the output of the last layer of the model.
+        cache_params (`Mamba2Cache`):
+            The state of the model at the last time step. Can be used in a forward method with the next `input_ids` to
+            avoid providing the old `input_ids`.
+
+            Includes both the State space model state matrices after the selective scan, and the Convolutional states
+        hidden_states (`tuple(torch.FloatTensor)`, *optional*,
+            returned when `output_hidden_states=True` is passed or when `config.output_hidden_states=True`):
+            Tuple of `torch.FloatTensor` (one for the output of the embeddings, if the model has an embedding layer, +
+            one for the output of each layer) of shape `(batch_size, sequence_length, hidden_size)`.
+
+            Hidden-states of the model at the output of each layer plus the optional initial embedding outputs.
+    """
+
+    last_hidden_state: Optional[torch.FloatTensor] = None
+    cache_params: Optional[Mamba2Cache] = None
+    hidden_states: Optional[Tuple[torch.FloatTensor]] = None
+
+
+@dataclass
+# Copied from transformers.models.mamba.modeling_mamba.MambaCausalLMOutput with Mamba->Mamba2
+class Mamba2CausalLMOutput(ModelOutput):
+    """
+    Base class for causal language model (or autoregressive) outputs.
+
+    Args:
+        loss (`torch.FloatTensor` of shape `(1,)`, *optional*, returned when `labels` is provided):
+            Language modeling loss (for next-token prediction).
+        logits (`torch.FloatTensor` of shape `(batch_size, sequence_length, config.vocab_size)`):
+            Prediction scores of the language modeling head (scores for each vocabulary token before SoftMax).
+        cache_params (`Mamba2Cache`):
+            The state of the model at the last time step. Can be used in a forward method with the next `input_ids` to
+            avoid providing the old `input_ids`.
+
+            Includes both the State space model state matrices after the selective scan, and the Convolutional states
+        hidden_states (`tuple(torch.FloatTensor)`, *optional*,
+            returned when `output_hidden_states=True` is passed or when `config.output_hidden_states=True`):
+            Tuple of `torch.FloatTensor` (one for the output of the embeddings, if the model has an embedding layer, +
+            one for the output of each layer) of shape `(batch_size, sequence_length, hidden_size)`.
+
+            Hidden-states of the model at the output of each layer plus the optional initial embedding outputs.
+    """
+
+    loss: Optional[torch.FloatTensor] = None
+    logits: Optional[torch.FloatTensor] = None
+    cache_params: Optional[Mamba2Cache] = None
+    hidden_states: Optional[Tuple[torch.FloatTensor]] = None
+
+
+class Mamba2Model(Mamba2PreTrainedModel):
+    def __init__(self, config):
+        super().__init__(config)
+
+        self.embeddings = nn.Embedding(config.vocab_size, config.hidden_size)
+        self.layers = nn.ModuleList([Mamba2Block(config, layer_idx=idx) for idx in range(config.num_hidden_layers)])
+
+        self.gradient_checkpointing = False
+        self.norm_f = RMSNorm(config.hidden_size, eps=config.layer_norm_epsilon)
+        # Initialize weights and apply final processing
+        self._register_load_state_dict_pre_hook(self.load_hook)
+        self.post_init()
+
+    def load_hook(self, state_dict, prefix, *args):
+        for k in state_dict:
+            if "embedding." in k:
+                state_dict[k.replace("embedding.", "embeddings.")] = state_dict.pop(k)
+                break
+
+    def get_input_embeddings(self):
+        return self.embeddings
+
+    def set_input_embeddings(self, new_embeddings):
+        self.embeddings = new_embeddings
+
+    def forward(
+        self,
+        input_ids: Optional[torch.LongTensor] = None,
+        inputs_embeds: Optional[torch.LongTensor] = None,
+        cache_params: Optional[Mamba2Cache] = None,
+        use_cache: Optional[bool] = None,
+        output_hidden_states: Optional[bool] = None,
+        return_dict: Optional[bool] = None,
+        cache_position: Optional[torch.LongTensor] = None,
+        attention_mask: Optional[torch.Tensor] = None,
+        **kwargs,
+    ) -> Union[Tuple, Mamba2Output]:
+        output_hidden_states = (
+            output_hidden_states if output_hidden_states is not None else self.config.output_hidden_states
+        )
+        use_cache = use_cache if use_cache is not None else (self.config.use_cache if not self.training else False)
+        return_dict = return_dict if return_dict is not None else self.config.use_return_dict
+
+        if (input_ids is None) ^ (inputs_embeds is not None):  # ^ is python for xor
+            raise ValueError("You must specify exactly one of input_ids or inputs_embeds")
+
+        if inputs_embeds is None:
+            inputs_embeds = self.embeddings(input_ids)
+
+        if self.gradient_checkpointing and self.training and use_cache:
+            use_cache = False
+
+        if use_cache:
+            if cache_params is None:
+                cache_params = Mamba2Cache(
+                    self.config, inputs_embeds.size(0), device=inputs_embeds.device, dtype=inputs_embeds.dtype
+                )
+                cache_position = torch.arange(0, self.config.conv_kernel, device=inputs_embeds.device)
+            elif cache_position is None:
+                # cases when we do manual forward instead of using `model.generate` which will initiate
+                # `cache_position` and makes sure it is not None, throw error here instead of doing some
+                # hack to conjecture the current cache position
+                raise ValueError(
+                    "You have to specify the `cache_position` manually when `use_cache=True` and `cache_params` is passed, "
+                    "you don't have to pass a `cache_params` if you are in prefilling stage because in that case it will "
+                    "be initialized for you automatically"
+                )
+        else:
+            cache_params = None
+
+        hidden_states = inputs_embeds
+        all_hidden_states = () if output_hidden_states else None
+        for mixer_block in self.layers:
+            if self.gradient_checkpointing and self.training:
+                hidden_states = self._gradient_checkpointing_func(
+                    mixer_block.__call__,
+                    hidden_states,
+                    cache_params,
+                    cache_position,
+                    attention_mask,
+                )
+            else:
+                hidden_states = mixer_block(
+                    hidden_states,
+                    cache_params=cache_params,
+                    cache_position=cache_position,
+                    attention_mask=attention_mask,
+                )
+
+            if output_hidden_states:
+                all_hidden_states = all_hidden_states + (hidden_states,)
+
+        hidden_states = self.norm_f(hidden_states)
+
+        if output_hidden_states:
+            all_hidden_states = all_hidden_states + (hidden_states,)
+
+        if not return_dict:
+            return tuple(v for v in [hidden_states, cache_params, all_hidden_states] if v is not None)
+
+        return Mamba2Output(
+            last_hidden_state=hidden_states,
+            cache_params=cache_params if use_cache else None,
+            hidden_states=all_hidden_states,
+        )
+
+
+class Mamba2ForCausalLM(Mamba2PreTrainedModel):
+    _tied_weights_keys = []
+
+    def __init__(self, config):
+        super().__init__(config)
+        self.backbone = Mamba2Model(config)
+        self.lm_head = nn.Linear(config.hidden_size, config.vocab_size, bias=False)
+        self.criterion = None
+
+        # Initialize weights and apply final processing
+        self.post_init()
+
+    def get_output_embeddings(self):
+        return self.lm_head
+
+    def set_output_embeddings(self, new_embeddings):
+        self.lm_head = new_embeddings
+
+    def get_input_embeddings(self):
+        return self.backbone.get_input_embeddings()
+
+    def set_input_embeddings(self, new_embeddings):
+        return self.backbone.set_input_embeddings(new_embeddings)
+
+    @deprecate_kwarg("num_logits_to_keep", version="4.50", new_name="logits_to_keep")
+    def prepare_inputs_for_generation(
+        self,
+        input_ids,
+        inputs_embeds=None,
+        use_cache=None,
+        cache_params: Optional[Mamba2Cache] = None,
+        cache_position: Optional[torch.LongTensor] = None,
+        attention_mask: Optional[torch.Tensor] = None,
+        logits_to_keep: Optional[int] = None,
+        **kwargs,
+    ):
+        if use_cache:
+            # `cache_position` should have been initialized in `generate`
+            if cache_position is None:
+                raise ValueError(
+                    "`cache_position` should not be None as it should have been initialized in "
+                    "`model.generate`, you are responsible for passing in a valid `cache_position` if "
+                    "you are calling `prepare_inputs_for_generation` directly with `use_cache=True`"
+                )
+            if cache_position[0] > 0:
+                input_ids = input_ids[:, -1][..., None]
+
+                if attention_mask is not None:
+                    attention_mask = None
+            else:
+                # we initialize the `cache_position` to full size of `conv_states` at prefill stage
+                # considering padding will be applied when input length is shorter, and truncation
+                # will be applied when it is longer, so it will be equivalent to always have it match
+                # the length of `cache_params.conv_states`, which is `config.conv_kernel`
+                cache_position = torch.arange(0, self.config.conv_kernel, device=input_ids.device)
+
+        if inputs_embeds is not None and cache_params is None:
+            model_inputs = {"inputs_embeds": inputs_embeds}
+        else:
+            model_inputs = {"input_ids": input_ids}
+
+        if logits_to_keep is not None:
+            model_inputs['logits_to_keep'] = logits_to_keep
+
+        model_inputs.update({
+            'attention_mask': attention_mask,
+            'cache_params': cache_params,
+            'use_cache': use_cache,
+            'cache_position': cache_position,
+            'logits_to_keep': logits_to_keep
+        })
+        return model_inputs
+
+    @deprecate_kwarg("num_logits_to_keep", version="4.50", new_name="logits_to_keep")
+    def forward(
+        self,
+        input_ids: Optional[torch.LongTensor] = None,
+        inputs_embeds: Optional[torch.FloatTensor] = None,
+        cache_params: Optional[Mamba2Cache] = None,
+        labels: Optional[torch.LongTensor] = None,
+        output_hidden_states: Optional[bool] = None,
+        return_dict: Optional[bool] = None,
+        use_cache: Optional[bool] = None,
+        cache_position: Optional[torch.Tensor] = None,
+        attention_mask: Optional[torch.Tensor] = None,
+        logits_to_keep: Optional[int] = 0,
+        **kwargs,  # for now we need this for generation
+    ) -> Union[Tuple, Mamba2CausalLMOutput]:
+        r"""
+        labels (`torch.LongTensor` of shape `(batch_size, sequence_length)`, *optional*):
+            Labels for language modeling. Note that the labels **are shifted** inside the model, i.e. you can set
+            `labels = input_ids` Indices are selected in `[-100, 0, ..., config.vocab_size]` All labels set to `-100`
+            are ignored (masked), the loss is only computed for labels in `[0, ..., config.vocab_size]`
+        """
+        return_dict = return_dict if return_dict is not None else self.config.use_return_dict
+
+        outputs = self.backbone(
+            input_ids,
+            cache_params=cache_params,
+            inputs_embeds=inputs_embeds,
+            output_hidden_states=output_hidden_states,
+            return_dict=return_dict,
+            use_cache=use_cache,
+            cache_position=cache_position,
+            attention_mask=attention_mask,
+        )
+        hidden_states = outputs[0]
+        fuse_linear_and_cross_entropy = self.config.fuse_cross_entropy and self.training
+
+        loss, logits = None, None
+        if not fuse_linear_and_cross_entropy or labels is None:
+            logits = self.lm_head(hidden_states if logits_to_keep is None else hidden_states[:, -logits_to_keep:])
+        if labels is not None:
+            if getattr(self, 'criterion', None) is None:
+                if fuse_linear_and_cross_entropy:
+                    criterion = FusedLinearCrossEntropyLoss()
+                elif self.config.fuse_cross_entropy:
+                    criterion = FusedCrossEntropyLoss(inplace_backward=True)
+                else:
+                    criterion = nn.CrossEntropyLoss()
+            else:
+                criterion = self.criterion
+            labels = labels.to(hidden_states.device)
+            labels = torch.cat((labels[..., 1:], torch.full_like(labels[:, :1], criterion.ignore_index)), 1)
+            if fuse_linear_and_cross_entropy:
+                loss = criterion(hidden_states, labels, self.lm_head.weight, self.lm_head.bias)
+            else:
+                loss = criterion(logits.view(labels.numel(), -1), labels.view(-1))
+
+        if not return_dict:
+            output = (logits,) + outputs[1:]
+            return (loss,) + output if loss is not None else output
+
+        return Mamba2CausalLMOutput(
+            loss=loss,
+            logits=logits,
+            cache_params=outputs.cache_params,
+            hidden_states=outputs.hidden_states,
+        )

fla/models/nsa/configuration_nsa.py

@@ -0,0 +1,75 @@
+# -*- coding: utf-8 -*-
+
+from typing import Optional
+
+from transformers.configuration_utils import PretrainedConfig
+
+
+class NSAConfig(PretrainedConfig):
+
+    model_type = 'nsa'
+    keys_to_ignore_at_inference = ['past_key_values']
+
+    def __init__(
+        self,
+        hidden_size: int = 2048,
+        num_hidden_layers: int = 24,
+        num_heads: int = 64,
+        num_kv_heads: int = 4,
+        head_dim: int = 32,
+        qkv_bias: bool = False,
+        block_size: int = 64,
+        block_counts: Optional[int] = 16,
+        window_size: Optional[int] = 512,
+        rope_theta: Optional[float] = 10000.,
+        max_position_embeddings: int = 2048,
+        hidden_ratio: Optional[int] = 4,
+        intermediate_size: Optional[int] = None,
+        hidden_act: str = "swish",
+        initializer_range: float = 0.006,
+        elementwise_affine: Optional[bool] = True,
+        norm_eps: float = 1e-6,
+        use_cache: bool = True,
+        pad_token_id: int = None,
+        bos_token_id: int = 1,
+        eos_token_id: int = 2,
+        tie_word_embeddings: bool = False,
+        fuse_norm: bool = True,
+        fuse_swiglu: bool = True,
+        fuse_cross_entropy: bool = True,
+        vocab_size: int = 32000,
+        **kwargs,
+    ):
+        self.hidden_size = hidden_size
+        self.num_hidden_layers = num_hidden_layers
+        self.num_heads = num_heads
+        self.num_kv_heads = num_kv_heads
+        self.head_dim = head_dim
+        self.qkv_bias = qkv_bias
+        self.block_size = block_size
+        self.block_counts = block_counts
+        self.window_size = window_size
+        self.rope_theta = rope_theta
+        self.max_position_embeddings = max_position_embeddings
+
+        self.hidden_ratio = hidden_ratio
+        self.intermediate_size = intermediate_size
+        self.hidden_act = hidden_act
+
+        self.initializer_range = initializer_range
+        self.elementwise_affine = elementwise_affine
+        self.norm_eps = norm_eps
+        self.use_cache = use_cache
+
+        self.fuse_norm = fuse_norm
+        self.fuse_swiglu = fuse_swiglu
+        self.fuse_cross_entropy = fuse_cross_entropy
+        self.vocab_size = vocab_size
+
+        super().__init__(
+            pad_token_id=pad_token_id,
+            bos_token_id=bos_token_id,
+            eos_token_id=eos_token_id,
+            tie_word_embeddings=tie_word_embeddings,
+            **kwargs,
+        )

fla/models/rwkv6/configuration_rwkv6.py

@@ -0,0 +1,82 @@
+# -*- coding: utf-8 -*-
+
+from typing import Dict, Optional
+
+from transformers.configuration_utils import PretrainedConfig
+
+
+class RWKV6Config(PretrainedConfig):
+
+    model_type = 'rwkv6'
+    keys_to_ignore_at_inference = ['past_key_values']
+
+    def __init__(
+        self,
+        attn_mode: str = "chunk",
+        hidden_size: int = 2048,
+        expand_k: int = 0.5,
+        expand_v: int = 1,
+        hidden_ratio: Optional[int] = 3.5,
+        intermediate_size: Optional[int] = None,
+        num_hidden_layers: int = 24,
+        num_heads: int = 4,
+        proj_low_rank_dim: int = 32,
+        gate_low_rank_dim: int = 64,
+        hidden_act: str = "sqrelu",
+        max_position_embeddings: int = 2048,
+        norm_first: bool = True,
+        norm_bias: bool = True,
+        norm_eps: float = 1e-5,
+        attn: Optional[Dict] = None,
+        use_cache: bool = True,
+        pad_token_id: int = None,
+        bos_token_id: int = 1,
+        eos_token_id: int = 2,
+        tie_word_embeddings: bool = False,
+        initializer_range: float = 0.006,
+        fuse_norm: bool = True,
+        fuse_cross_entropy: bool = True,
+        vocab_size: int = 32000,
+        **kwargs
+    ):
+        self.attn_mode = attn_mode
+        self.hidden_size = hidden_size
+        self.expand_k = expand_k
+        self.expand_v = expand_v
+        self.hidden_ratio = hidden_ratio
+        self.intermediate_size = intermediate_size
+        self.norm_first = norm_first
+        self.num_hidden_layers = num_hidden_layers
+        self.num_heads = num_heads
+        self.proj_low_rank_dim = proj_low_rank_dim
+        self.gate_low_rank_dim = gate_low_rank_dim
+        self.hidden_act = hidden_act
+        self.max_position_embeddings = max_position_embeddings
+        self.norm_bias = norm_bias
+        self.norm_eps = norm_eps
+        self.attn = attn
+        self.use_cache = use_cache
+        self.initializer_range = initializer_range
+        self.fuse_norm = fuse_norm
+        self.fuse_cross_entropy = fuse_cross_entropy
+        self.vocab_size = vocab_size
+
+        if attn is not None:
+            if not isinstance(attn, Dict):
+                raise ValueError("attn must be a dictionary")
+            if 'layers' not in attn:
+                raise ValueError("Layer indices must be provided to initialize hybrid attention layers")
+            if 'num_heads' not in attn:
+                raise ValueError("Number of heads must be provided to initialize hybrid attention layers")
+            attn['num_kv_heads'] = attn.get('num_kv_heads', attn['num_heads'])
+            attn['qkv_bias'] = attn.get('qkv_bias', False)
+            attn['window_size'] = attn.get('window_size', None)
+            attn['rope_theta'] = attn.get('rope_theta', 10000.)
+
+        super().__init__(
+            pad_token_id=pad_token_id,
+            bos_token_id=bos_token_id,
+            eos_token_id=eos_token_id,
+            tie_word_embeddings=tie_word_embeddings,
+            **kwargs,
+        )

fla/models/rwkv6/modeling_rwkv6.py

@@ -0,0 +1,480 @@
+# -*- coding: utf-8 -*-
+
+from __future__ import annotations
+
+import math
+import warnings
+from typing import TYPE_CHECKING, Dict, Optional, Tuple, Union
+
+import torch
+import torch.nn as nn
+import torch.utils.checkpoint
+from transformers.generation import GenerationMixin
+from transformers.modeling_outputs import BaseModelOutputWithPast, CausalLMOutputWithPast
+from transformers.modeling_utils import PreTrainedModel
+from transformers.utils import logging
+from transformers.utils.deprecation import deprecate_kwarg
+
+from fla.layers.attn import Attention
+from fla.layers.rwkv6 import LerpLinear, RWKV6Attention
+from fla.models.rwkv6.configuration_rwkv6 import RWKV6Config
+from fla.models.utils import Cache
+from fla.modules import FusedCrossEntropyLoss, FusedLinearCrossEntropyLoss, LayerNorm
+from fla.modules.activations import ACT2FN
+
+if TYPE_CHECKING:
+    from transformers.processing_utils import Unpack
+
+logger = logging.get_logger(__name__)
+
+
+class RWKV6FeedForward(nn.Module):
+
+    def __init__(
+        self,
+        hidden_size: int,
+        hidden_ratio: Optional[int] = None,
+        intermediate_size: Optional[int] = None,
+        hidden_act: str = 'sqrelu',
+        layer_idx: int = None
+    ) -> RWKV6FeedForward:
+        super().__init__()
+
+        self.hidden_size = hidden_size
+        if hidden_ratio is None:
+            hidden_ratio = 3.5
+        if intermediate_size is None:
+            intermediate_size = int(hidden_size * hidden_ratio)
+            intermediate_size = 32 * ((intermediate_size + 32 - 1) // 32)
+        self.hidden_ratio = hidden_ratio
+        self.intermediate_size = intermediate_size
+
+        self.time_shift = nn.ZeroPad2d((0, 0, 1, -1))
+
+        self.key = LerpLinear(hidden_size, intermediate_size)
+        self.value = nn.Linear(intermediate_size, hidden_size, bias=False)
+        self.receptance = LerpLinear(hidden_size, hidden_size)
+        self.act_fn = ACT2FN[hidden_act]
+
+        self.layer_idx = layer_idx
+
+    def forward(
+        self,
+        x: torch.Tensor,
+        attention_mask: Optional[torch.Tensor] = None,
+        state: Optional[Cache] = None
+    ) -> torch.Tensor:
+        if attention_mask is not None:
+            x = x.mul_(attention_mask[:, -x.shape[-2]:, None])
+        if x.shape[1] == 1 and state is not None and state[self.layer_idx]['ffn_state'] is not None:
+            shifted = state[self.layer_idx]['ffn_state'].unsqueeze(1)
+        else:
+            shifted = self.time_shift(x)
+            if state is not None and state[self.layer_idx]['ffn_state'] is not None:
+                shifted[:, 0] = state[self.layer_idx]['ffn_state']
+        delta = shifted - x
+        key = self.act_fn(self.key(x, delta))
+        value = self.value(key)
+        receptance = self.receptance(x, delta)
+
+        if state is not None:
+            # no need to update the offset twice
+            state.update(ffn_state=x[:, -1], layer_idx=self.layer_idx, offset=0)
+        return receptance.sigmoid() * value, state
+
+
+class RWKV6Block(nn.Module):
+    def __init__(self, config: RWKV6Config, layer_idx: int):
+        super().__init__()
+
+        self.config = config
+        self.layer_idx = layer_idx
+
+        if config.norm_first and layer_idx == 0:
+            self.pre_norm = (LayerNorm if config.fuse_norm else nn.LayerNorm)(
+                config.hidden_size,
+                bias=config.norm_bias,
+                eps=config.norm_eps
+            )
+        self.attn_norm = (LayerNorm if config.fuse_norm else nn.LayerNorm)(
+            config.hidden_size,
+            bias=config.norm_bias,
+            eps=config.norm_eps
+        )
+        if config.attn is not None and layer_idx in config.attn['layers']:
+            self.attn = Attention(
+                hidden_size=config.hidden_size,
+                num_heads=config.attn['num_heads'],
+                num_kv_heads=config.attn['num_kv_heads'],
+                qkv_bias=config.attn['qkv_bias'],
+                window_size=config.attn['window_size'],
+                rope_theta=config.attn['rope_theta'],
+                max_position_embeddings=config.max_position_embeddings,
+                layer_idx=layer_idx
+            )
+        else:
+            self.attn = RWKV6Attention(
+                mode=config.attn_mode,
+                hidden_size=config.hidden_size,
+                expand_k=config.expand_k,
+                expand_v=config.expand_v,
+                num_heads=config.num_heads,
+                proj_low_rank_dim=config.proj_low_rank_dim,
+                gate_low_rank_dim=config.gate_low_rank_dim,
+                norm_eps=config.norm_eps,
+                fuse_norm=config.fuse_norm,
+                layer_idx=layer_idx
+            )
+        self.ffn_norm = (LayerNorm if config.fuse_norm else nn.LayerNorm)(
+            config.hidden_size,
+            bias=config.norm_bias,
+            eps=config.norm_eps
+        )
+        self.ffn = RWKV6FeedForward(
+            hidden_size=config.hidden_size,
+            hidden_ratio=config.hidden_ratio,
+            intermediate_size=config.intermediate_size,
+            hidden_act=config.hidden_act,
+            layer_idx=layer_idx
+        )
+
+    def forward(
+        self,
+        hidden_states: torch.Tensor,
+        attention_mask: Optional[torch.Tensor] = None,
+        past_key_values: Optional[Cache] = None,
+        use_cache: Optional[bool] = False,
+        output_attentions: Optional[bool] = False,
+        **kwargs,
+    ) -> Tuple[torch.FloatTensor, Optional[Tuple[torch.FloatTensor, torch.FloatTensor]]]:
+        residual = self.pre_norm(hidden_states) if hasattr(self, 'pre_norm') else hidden_states
+        hidden_states = self.attn_norm(residual)
+        hidden_states, attentions, past_key_values = self.attn(
+            hidden_states=hidden_states,
+            attention_mask=attention_mask,
+            past_key_values=past_key_values,
+            use_cache=use_cache,
+            output_attentions=output_attentions,
+            **kwargs
+        )
+        if self.config.fuse_norm:
+            hidden_states, residual = self.ffn_norm(hidden_states, residual, True)
+        else:
+            hidden_states = residual + hidden_states
+            residual = hidden_states
+            hidden_states = self.ffn_norm(hidden_states)
+        hidden_states, past_key_values = self.ffn(hidden_states, attention_mask, past_key_values)
+        hidden_states = residual + hidden_states
+
+        outputs = (hidden_states, attentions, past_key_values)
+
+        return outputs
+
+
+class RWKV6PreTrainedModel(PreTrainedModel):
+
+    config_class = RWKV6Config
+    base_model_prefix = 'model'
+    supports_gradient_checkpointing = True
+    _no_split_modules = ['RWKV6Block']
+    _supports_cache_class = True
+
+    def __init__(self, *inputs, **kwargs):
+        super().__init__(*inputs, **kwargs)
+
+    def _init_weights(
+        self,
+        module: nn.Module,
+        rescale_prenorm_residual: bool = True,
+        num_residuals_per_layer: int = 2,
+    ):
+        if isinstance(module, (nn.Linear, nn.Conv1d)):
+            # Slightly different from the TF version which uses truncated_normal for initialization
+            # cf https://github.com/pytorch/pytorch/pull/5617
+            nn.init.normal_(module.weight, mean=0.0, std=self.config.initializer_range)
+            if module.bias is not None:
+                nn.init.zeros_(module.bias)
+        elif isinstance(module, nn.Parameter):
+            nn.init.normal_(module, mean=0.0, std=self.config.initializer_range)
+        elif isinstance(module, nn.Embedding):
+            nn.init.normal_(module.weight, mean=0.0, std=self.config.initializer_range)
+        elif hasattr(module, 'reset_parameters'):
+            module.reset_parameters()
+
+        if rescale_prenorm_residual:
+            # Reinitialize selected weights subject to the OpenAI GPT-2 Paper Scheme:
+            #   > A modified initialization which accounts for the accumulation on the residual path with model depth. Scale
+            #   > the weights of residual layers at initialization by a factor of 1/√N where N is the # of residual layers.
+            #   >   -- GPT-2 :: https://openai.com/blog/better-language-models/
+            #
+            # Reference (Megatron-LM): https://github.com/NVIDIA/Megatron-LM/blob/main/megatron/model/gpt_model.py
+            p = None
+            if hasattr(module, 'o_proj'):
+                p = module.o_proj.weight
+            elif hasattr(module, 'down_proj'):
+                p = module.down_proj.weight
+            if p is not None:
+                # Special Scaled Initialization --> There are 2 Layer Norms per Transformer Block
+                # Following Pytorch init, except scale by 1/sqrt(2 * n_layer)
+                # We need to reinit p since this code could be called multiple times
+                # Having just p *= scale would repeatedly scale it down
+                nn.init.kaiming_uniform_(p, a=math.sqrt(5))
+                with torch.no_grad():
+                    p /= math.sqrt(num_residuals_per_layer * self.config.num_hidden_layers)
+
+
+class RWKV6Model(RWKV6PreTrainedModel):
+
+    def __init__(self, config: RWKV6Config):
+        super().__init__(config)
+        self.padding_idx = config.pad_token_id
+        self.vocab_size = config.vocab_size
+
+        self.embeddings = nn.Embedding(config.vocab_size, config.hidden_size, self.padding_idx)
+        self.layers = nn.ModuleList([RWKV6Block(config, layer_idx) for layer_idx in range(config.num_hidden_layers)])
+        self.norm = (LayerNorm if config.fuse_norm else nn.LayerNorm)(
+            config.hidden_size,
+            bias=config.norm_bias,
+            eps=config.norm_eps
+        )
+
+        self.gradient_checkpointing = False
+
+        self.post_init()
+
+    def get_input_embeddings(self):
+        return self.embeddings
+
+    def set_input_embeddings(self, value):
+        self.embeddings = value
+
+    def forward(
+        self,
+        input_ids: Optional[torch.LongTensor] = None,
+        attention_mask: Optional[torch.Tensor] = None,  # noqa
+        inputs_embeds: Optional[torch.FloatTensor] = None,
+        past_key_values: Optional[Cache] = None,
+        use_cache: Optional[bool] = None,
+        output_attentions: Optional[bool] = None,
+        output_hidden_states: Optional[bool] = None,
+        return_dict: Optional[bool] = None,
+        **kwargs: Unpack[Dict]
+    ) -> Union[Tuple, BaseModelOutputWithPast]:
+        if output_attentions:
+            warnings.warn("`RWKV6Model` does not `output_attentions` now, setting it to `False`.")
+            output_attentions = False
+        output_attentions = output_attentions if output_attentions is not None else self.config.output_attentions
+        output_hidden_states = output_hidden_states if output_hidden_states is not None else self.config.output_hidden_states
+        use_cache = use_cache if use_cache is not None else (self.config.use_cache if not self.training else False)
+        return_dict = return_dict if return_dict is not None else self.config.use_return_dict
+
+        # retrieve input_ids and inputs_embeds
+        if input_ids is not None and inputs_embeds is not None:
+            raise ValueError("You cannot specify both input_ids and inputs_embeds at the same time")
+        if input_ids is None and inputs_embeds is None:
+            raise ValueError("You have to specify either input_ids or inputs_embeds")
+
+        if inputs_embeds is None:
+            inputs_embeds = self.embeddings(input_ids)
+        hidden_states = inputs_embeds
+
+        if use_cache and not isinstance(past_key_values, Cache):
+            past_key_values = Cache.from_legacy_cache(past_key_values)
+
+        if self.gradient_checkpointing and self.training and use_cache:
+            logger.warning_once("`use_cache=True` is incompatible with gradient checkpointing. Setting `use_cache=False`...")
+            use_cache = False
+
+        all_hidden_states = () if output_hidden_states else None
+        all_attns = () if output_attentions else None
+        for layer in self.layers:
+            if output_hidden_states:
+                all_hidden_states += (hidden_states,)
+
+            if self.gradient_checkpointing and self.training:
+                hidden_states, attentions, past_key_values = self._gradient_checkpointing_func(
+                    layer.__call__,
+                    hidden_states,
+                    attention_mask,
+                    past_key_values,
+                    use_cache,
+                    output_attentions,
+                    **kwargs
+                )
+            else:
+                hidden_states, attentions, past_key_values = layer(
+                    hidden_states,
+                    attention_mask=attention_mask,
+                    past_key_values=past_key_values,
+                    use_cache=use_cache,
+                    output_attentions=output_attentions,
+                    **kwargs
+                )
+
+            if output_attentions:
+                all_attns += (attentions,)
+
+        hidden_states = self.norm(hidden_states)
+
+        # add hidden states from the last decoder layer
+        if output_hidden_states:
+            all_hidden_states += (hidden_states,)
+
+        if not return_dict:
+            return tuple(i for i in [hidden_states, past_key_values, all_hidden_states, all_attns] if i is not None)
+        return BaseModelOutputWithPast(
+            last_hidden_state=hidden_states,
+            past_key_values=past_key_values,
+            hidden_states=all_hidden_states,
+            attentions=all_attns
+        )
+
+
+class RWKV6ForCausalLM(RWKV6PreTrainedModel, GenerationMixin):
+
+    _tied_weights_keys = ["lm_head.weight"]
+
+    def __init__(self, config):
+        super().__init__(config)
+        self.model = RWKV6Model(config)
+        self.vocab_size = config.vocab_size
+        self.lm_head = nn.Linear(config.hidden_size, config.vocab_size, bias=False)
+        self.criterion = None
+
+        # Initialize weights and apply final processing
+        self.post_init()
+
+    def get_input_embeddings(self):
+        return self.model.embeddings
+
+    def set_input_embeddings(self, value):
+        self.model.embeddings = value
+
+    def get_output_embeddings(self):
+        return self.lm_head
+
+    def set_output_embeddings(self, new_embeddings):
+        self.lm_head = new_embeddings
+
+    def set_decoder(self, decoder):
+        self.model = decoder
+
+    def get_decoder(self):
+        return self.model
+
+    def generate(self, *args, **kwargs):
+        try:
+            return super().generate(*args, **kwargs)
+        except AttributeError as exception:
+            if 'past_key_values' in str(exception):
+                raise AttributeError(
+                    f"You tried to call `generate` with a decoding strategy that manipulates `past_key_values`, "
+                    f"which is not supported for {self.__class__.__name__}. "
+                    f"Try another generation strategy instead. "
+                    f"For the available generation strategies, check this doc: "
+                    f"https://huggingface.co/docs/transformers/en/generation_strategies#decoding-strategies"
+                )
+            else:
+                raise exception
+
+    @deprecate_kwarg("num_logits_to_keep", version="4.50", new_name="logits_to_keep")
+    def prepare_inputs_for_generation(
+        self,
+        input_ids: torch.LongTensor = None,
+        past_key_values: Optional[Cache] = None,
+        attention_mask: Optional[torch.Tensor] = None,
+        inputs_embeds: Optional[torch.Tensor] = None,
+        use_cache: bool = True,
+        logits_to_keep: Optional[int] = None,
+        **kwargs
+    ):
+        # only last token for `inputs_ids` if the `past_key_values` is not empty.
+        if past_key_values is not None and len(past_key_values) > 0:
+            input_ids = input_ids[:, -1:]
+        # if `inputs_embeds` are passed, we only want to use them in the 1st generation step
+        if inputs_embeds is not None and len(past_key_values) == 0:
+            model_inputs = {'inputs_embeds': inputs_embeds}
+        else:
+            # The `contiguous()` here is necessary to have a static stride during decoding. torchdynamo otherwise
+            # recompiles graphs as the stride of the inputs is a guard.
+            # Ref: https://github.com/huggingface/transformers/pull/29114
+            # TODO: use `next_tokens` directly instead.
+            model_inputs = {'input_ids': input_ids.contiguous()}
+
+        if logits_to_keep is not None:
+            model_inputs['logits_to_keep'] = logits_to_keep
+
+        model_inputs.update({
+            'past_key_values': past_key_values,
+            'use_cache': use_cache,
+            'attention_mask': attention_mask,
+        })
+        return model_inputs
+
+    @deprecate_kwarg("num_logits_to_keep", version="4.50", new_name="logits_to_keep")
+    def forward(
+        self,
+        input_ids: torch.LongTensor = None,
+        attention_mask: Optional[torch.Tensor] = None,
+        inputs_embeds: Optional[torch.Tensor] = None,
+        past_key_values: Optional[Cache] = None,
+        labels: Optional[torch.LongTensor] = None,
+        use_cache: Optional[bool] = None,
+        output_attentions: Optional[bool] = None,
+        output_hidden_states: Optional[bool] = None,
+        return_dict: Optional[bool] = None,
+        logits_to_keep: Optional[int] = 0,
+        **kwargs: Unpack[Dict]
+    ) -> Union[Tuple, CausalLMOutputWithPast]:
+        output_attentions = output_attentions if output_attentions is not None else self.config.output_attentions
+        output_hidden_states = (
+            output_hidden_states if output_hidden_states is not None else self.config.output_hidden_states
+        )
+        return_dict = return_dict if return_dict is not None else self.config.use_return_dict
+
+        outputs = self.model(
+            input_ids=input_ids,
+            attention_mask=attention_mask,
+            inputs_embeds=inputs_embeds,
+            past_key_values=past_key_values,
+            use_cache=use_cache,
+            output_attentions=output_attentions,
+            output_hidden_states=output_hidden_states,
+            return_dict=return_dict,
+            **kwargs
+        )
+
+        hidden_states = outputs[0]
+        fuse_linear_and_cross_entropy = self.config.fuse_cross_entropy and self.training
+
+        loss, logits = None, None
+        if not fuse_linear_and_cross_entropy or labels is None:
+            logits = self.lm_head(hidden_states if logits_to_keep is None else hidden_states[:, -logits_to_keep:])
+        if labels is not None:
+            if getattr(self, 'criterion', None) is None:
+                if fuse_linear_and_cross_entropy:
+                    criterion = FusedLinearCrossEntropyLoss()
+                elif self.config.fuse_cross_entropy:
+                    criterion = FusedCrossEntropyLoss(inplace_backward=True)
+                else:
+                    criterion = nn.CrossEntropyLoss()
+            else:
+                criterion = self.criterion
+            labels = labels.to(hidden_states.device)
+            labels = torch.cat((labels[..., 1:], torch.full_like(labels[:, :1], criterion.ignore_index)), 1)
+            if fuse_linear_and_cross_entropy:
+                loss = criterion(hidden_states, labels, self.lm_head.weight, self.lm_head.bias)
+            else:
+                loss = criterion(logits.view(labels.numel(), -1), labels.view(-1))
+
+        if not return_dict:
+            output = (logits,) + outputs[1:]
+            return (loss,) + output if loss is not None else output
+
+        return CausalLMOutputWithPast(
+            loss=loss,
+            logits=logits,
+            past_key_values=outputs.past_key_values,
+            hidden_states=outputs.hidden_states,
+            attentions=outputs.attentions,
+        )

fla/models/rwkv7/configuration_rwkv7.py

@@ -0,0 +1,105 @@
+# -*- coding: utf-8 -*-
+
+from typing import Dict, List, Optional, Union
+
+from transformers.configuration_utils import PretrainedConfig
+
+
+class RWKV7Config(PretrainedConfig):
+
+    model_type = 'rwkv7'
+    keys_to_ignore_at_inference = ['past_key_values']
+
+    def __init__(
+        self,
+        attn_mode: str = "chunk",
+        hidden_size: int = 2048,
+        hidden_ratio: Optional[int] = 4,
+        intermediate_size: Optional[int] = None,
+        num_hidden_layers: int = 24,
+        head_dim: Optional[int] = 64,
+        num_heads: Optional[int] = None,
+        decay_low_rank_dim: int = 64,
+        gate_low_rank_dim: int = 128,
+        a_low_rank_dim: int = 64,
+        v_low_rank_dim: int = 16,
+        hidden_act: str = "sqrelu",
+        max_position_embeddings: int = 2048,
+        norm_first: bool = True,
+        norm_bias: bool = True,
+        norm_eps: float = 1e-5,
+        attn: Optional[Dict] = None,
+        use_cache: bool = True,
+        pad_token_id: int = None,
+        bos_token_id: int = 1,
+        eos_token_id: int = 2,
+        tie_word_embeddings: bool = False,
+        initializer_range: float = 0.006,
+        fuse_norm: bool = True,
+        fuse_cross_entropy: bool = True,
+        vocab_size: int = 32000,
+        value_dim: Optional[Union[int, List[int]]] = None,
+        **kwargs
+    ):
+        self.attn_mode = attn_mode
+        self.hidden_size = hidden_size
+        self.hidden_ratio = hidden_ratio
+        self.intermediate_size = intermediate_size
+        self.norm_first = norm_first
+        self.num_hidden_layers = num_hidden_layers
+
+        if head_dim is None and num_heads is not None:
+            head_dim = int(hidden_size // num_heads)
+        elif head_dim is not None and num_heads is None:
+            num_heads = int(hidden_size // head_dim)
+
+        if value_dim is None:
+            value_dim = [hidden_size] * num_hidden_layers
+        elif isinstance(value_dim, int):
+            assert value_dim >= hidden_size, "value_dim must be greater than hidden_size"
+            assert value_dim % hidden_size == 0, "value_dim must be divisible by hidden_size"
+            value_dim = [value_dim] * num_hidden_layers
+        else:
+            assert len(value_dim) == num_hidden_layers, "value_dim must have the same length as num_hidden_layers"
+            for v in value_dim:
+                assert v >= hidden_size, "value_dim must be greater than hidden_size"
+                assert v % hidden_size == 0, "value_dim must be divisible by hidden_size"
+
+        self.head_dim = head_dim
+        self.num_heads = num_heads
+        self.value_dim = value_dim
+
+        self.decay_low_rank_dim = decay_low_rank_dim
+        self.gate_low_rank_dim = gate_low_rank_dim
+        self.a_low_rank_dim = a_low_rank_dim
+        self.v_low_rank_dim = v_low_rank_dim
+        self.hidden_act = hidden_act
+        self.max_position_embeddings = max_position_embeddings
+        self.norm_bias = norm_bias
+        self.norm_eps = norm_eps
+        self.attn = attn
+        self.use_cache = use_cache
+        self.initializer_range = initializer_range
+        self.fuse_norm = fuse_norm
+        self.fuse_cross_entropy = fuse_cross_entropy
+        self.vocab_size = vocab_size
+
+        if attn is not None:
+            if not isinstance(attn, Dict):
+                raise ValueError("attn must be a dictionary")
+            if 'layers' not in attn:
+                raise ValueError("Layer indices must be provided to initialize hybrid attention layers")
+            if 'num_heads' not in attn:
+                raise ValueError("Number of heads must be provided to initialize hybrid attention layers")
+            attn['num_kv_heads'] = attn.get('num_kv_heads', attn['num_heads'])
+            attn['qkv_bias'] = attn.get('qkv_bias', False)
+            attn['window_size'] = attn.get('window_size', None)
+            attn['rope_theta'] = attn.get('rope_theta', 10000.)
+
+        super().__init__(
+            pad_token_id=pad_token_id,
+            bos_token_id=bos_token_id,
+            eos_token_id=eos_token_id,
+            tie_word_embeddings=tie_word_embeddings,
+            **kwargs,
+        )

fla/models/rwkv7/modeling_rwkv7.py

@@ -0,0 +1,505 @@
+# -*- coding: utf-8 -*-
+
+from __future__ import annotations
+
+import math
+import warnings
+from typing import TYPE_CHECKING, Dict, Optional, Tuple, Union
+
+import torch
+import torch.nn as nn
+import torch.utils.checkpoint
+from transformers.generation import GenerationMixin
+from transformers.modeling_outputs import BaseModelOutputWithPast, CausalLMOutputWithPast
+from transformers.modeling_utils import PreTrainedModel
+from transformers.utils import logging
+from transformers.utils.deprecation import deprecate_kwarg
+
+from fla.layers.attn import Attention
+from fla.layers.rwkv7 import RWKV7Attention
+from fla.models.rwkv7.configuration_rwkv7 import RWKV7Config
+from fla.models.utils import Cache
+from fla.modules import FusedCrossEntropyLoss, FusedLinearCrossEntropyLoss, LayerNorm
+from fla.modules.activations import ACT2FN
+
+if TYPE_CHECKING:
+    from transformers.processing_utils import Unpack
+
+logger = logging.get_logger(__name__)
+
+
+class RWKV7FeedForward(nn.Module):
+
+    def __init__(
+        self,
+        hidden_size: int,
+        hidden_ratio: Optional[int] = None,
+        intermediate_size: Optional[int] = None,
+        hidden_act: str = 'sqrelu',
+        layer_idx: int = None
+    ) -> RWKV7FeedForward:
+        super().__init__()
+
+        self.hidden_size = hidden_size
+        if hidden_ratio is None:
+            hidden_ratio = 4
+        if intermediate_size is None:
+            intermediate_size = int(hidden_size * hidden_ratio)
+            intermediate_size = 32 * ((intermediate_size + 32 - 1) // 32)
+        self.hidden_ratio = hidden_ratio
+        self.intermediate_size = intermediate_size
+
+        self.time_shift = nn.ZeroPad2d((0, 0, 1, -1))
+
+        self.x_k = nn.Parameter(torch.zeros(hidden_size))
+
+        self.key = nn.Linear(hidden_size, intermediate_size, bias=False)
+        self.value = nn.Linear(intermediate_size, hidden_size, bias=False)
+        self.act_fn = ACT2FN[hidden_act]
+
+        self.layer_idx = layer_idx
+
+    def forward(
+        self,
+        x: torch.Tensor,
+        attention_mask: Optional[torch.Tensor] = None,
+        state: Optional[Cache] = None
+    ) -> torch.Tensor:
+        if attention_mask is not None:
+            x = x.mul(attention_mask[:, -x.shape[-2]:, None])
+        if x.shape[1] == 1 and state is not None and state[self.layer_idx]['ffn_state'] is not None:
+            shifted = state[self.layer_idx]['ffn_state'].unsqueeze(1)
+        else:
+            shifted = self.time_shift(x)
+            if state is not None and state[self.layer_idx]['ffn_state'] is not None:
+                shifted[:, 0] = state[self.layer_idx]['ffn_state'][-1]
+        if state is not None:
+            # no need to update the offset twice
+            state.update(ffn_state=x[:, -1], layer_idx=self.layer_idx, offset=0)
+        return self.value(self.act_fn(self.key(x.addcmul(shifted - x, self.x_k)))), state
+
+
+class RWKV7Block(nn.Module):
+
+    def __init__(
+        self,
+        config: RWKV7Config,
+        layer_idx: int
+    ) -> RWKV7Block:
+        super().__init__()
+
+        self.config = config
+        self.layer_idx = layer_idx
+
+        if config.norm_first and layer_idx == 0:
+            self.pre_norm = (LayerNorm if config.fuse_norm else nn.LayerNorm)(
+                config.hidden_size,
+                bias=config.norm_bias,
+                eps=config.norm_eps
+            )
+        self.attn_norm = (LayerNorm if config.fuse_norm else nn.LayerNorm)(
+            config.hidden_size,
+            bias=config.norm_bias,
+            eps=config.norm_eps
+        )
+        if config.attn is not None and layer_idx in config.attn['layers']:
+            self.attn = Attention(
+                hidden_size=config.hidden_size,
+                num_heads=config.attn['num_heads'],
+                num_kv_heads=config.attn['num_kv_heads'],
+                qkv_bias=config.attn['qkv_bias'],
+                window_size=config.attn['window_size'],
+                rope_theta=config.attn['rope_theta'],
+                max_position_embeddings=config.max_position_embeddings,
+                layer_idx=layer_idx
+            )
+        else:
+            self.attn = RWKV7Attention(
+                mode=config.attn_mode,
+                hidden_size=config.hidden_size,
+                head_dim=config.head_dim,
+                num_heads=config.num_heads,
+                decay_low_rank_dim=config.decay_low_rank_dim,
+                gate_low_rank_dim=config.gate_low_rank_dim,
+                a_low_rank_dim=config.a_low_rank_dim,
+                v_low_rank_dim=config.v_low_rank_dim,
+                norm_eps=config.norm_eps,
+                fuse_norm=config.fuse_norm,
+                layer_idx=layer_idx,
+                value_dim=config.value_dim[layer_idx]
+            )
+        self.ffn_norm = (LayerNorm if config.fuse_norm else nn.LayerNorm)(
+            config.hidden_size,
+            bias=config.norm_bias,
+            eps=config.norm_eps
+        )
+        self.ffn = RWKV7FeedForward(
+            hidden_size=config.hidden_size,
+            hidden_ratio=config.hidden_ratio,
+            intermediate_size=config.intermediate_size,
+            hidden_act=config.hidden_act,
+            layer_idx=layer_idx
+        )
+
+    def forward(
+        self,
+        hidden_states: torch.Tensor,
+        attention_mask: Optional[torch.Tensor] = None,
+        past_key_values: Optional[Cache] = None,
+        use_cache: Optional[bool] = False,
+        output_attentions: Optional[bool] = False,
+        v_first: torch.Tensor = None,
+        **kwargs,
+    ) -> Tuple[torch.FloatTensor, Optional[Tuple[torch.FloatTensor, torch.FloatTensor]]]:
+        residual = self.pre_norm(hidden_states) if hasattr(self, 'pre_norm') else hidden_states
+        hidden_states = self.attn_norm(residual)
+        hidden_states, attentions, past_key_values, v_first = self.attn(
+            hidden_states=hidden_states,
+            attention_mask=attention_mask,
+            past_key_values=past_key_values,
+            use_cache=use_cache,
+            output_attentions=output_attentions,
+            v_first=v_first,
+            **kwargs
+        )
+        if self.config.fuse_norm:
+            hidden_states, residual = self.ffn_norm(hidden_states, residual, True)
+        else:
+            hidden_states = residual + hidden_states
+            residual = hidden_states
+            hidden_states = self.ffn_norm(hidden_states)
+        hidden_states, past_key_values = self.ffn(hidden_states, attention_mask, past_key_values)
+        hidden_states = residual + hidden_states
+
+        outputs = (hidden_states, attentions, past_key_values, v_first)
+
+        return outputs
+
+
+class RWKV7PreTrainedModel(PreTrainedModel):
+
+    config_class = RWKV7Config
+    base_model_prefix = 'model'
+    supports_gradient_checkpointing = True
+    _no_split_modules = ['RWKV7Block']
+    _supports_cache_class = True
+    _skip_keys_device_placement = ["past_key_values"]
+
+    def __init__(self, *inputs, **kwargs):
+        super().__init__(*inputs, **kwargs)
+
+    def _init_weights(
+        self,
+        module: nn.Module,
+        rescale_prenorm_residual: bool = True,
+        num_residuals_per_layer: int = 2,
+    ):
+        warnings.warn(
+            "RWKV-7 employs a carefully designed initialization strategy tailored to its architecture. "
+            "The detailed initialization scheme is currently not implemented here but can be found in the "
+            "official code repository. We emphasize that using the recommended initialization is essential "
+            "for replicating the results in RWKV-7 paper. Deviations from the prescribed initialization "
+            "may lead to performance degradation.\n"
+            "Alternatively, please generate initial weights from the official RWKV code repository, and "
+            "convert the PyTorch checkpoint into FLA supported format."
+        )
+        if isinstance(module, (nn.Linear, nn.Conv1d)):
+            # Slightly different from the TF version which uses truncated_normal for initialization
+            # cf https://github.com/pytorch/pytorch/pull/5617
+            nn.init.normal_(module.weight, mean=0.0, std=self.config.initializer_range)
+            if module.bias is not None:
+                nn.init.zeros_(module.bias)
+        elif isinstance(module, nn.Parameter):
+            nn.init.normal_(module, mean=0.0, std=self.config.initializer_range)
+        elif isinstance(module, nn.Embedding):
+            nn.init.normal_(module.weight, mean=0.0, std=self.config.initializer_range)
+        elif hasattr(module, 'reset_parameters'):
+            module.reset_parameters()
+
+        if rescale_prenorm_residual:
+            # Reinitialize selected weights subject to the OpenAI GPT-2 Paper Scheme:
+            #   > A modified initialization which accounts for the accumulation on the residual path with model depth. Scale
+            #   > the weights of residual layers at initialization by a factor of 1/√N where N is the # of residual layers.
+            #   >   -- GPT-2 :: https://openai.com/blog/better-language-models/
+            #
+            # Reference (Megatron-LM): https://github.com/NVIDIA/Megatron-LM/blob/main/megatron/model/gpt_model.py
+            p = None
+            if hasattr(module, 'o_proj'):
+                p = module.o_proj.weight
+            elif hasattr(module, 'down_proj'):
+                p = module.down_proj.weight
+            if p is not None:
+                # Special Scaled Initialization --> There are 2 Layer Norms per Transformer Block
+                # Following Pytorch init, except scale by 1/sqrt(2 * n_layer)
+                # We need to reinit p since this code could be called multiple times
+                # Having just p *= scale would repeatedly scale it down
+                nn.init.kaiming_uniform_(p, a=math.sqrt(5))
+                with torch.no_grad():
+                    p /= math.sqrt(num_residuals_per_layer * self.config.num_hidden_layers)
+
+
+class RWKV7Model(RWKV7PreTrainedModel):
+
+    def __init__(self, config: RWKV7Config):
+        super().__init__(config)
+        self.padding_idx = config.pad_token_id
+        self.vocab_size = config.vocab_size
+
+        self.embeddings = nn.Embedding(config.vocab_size, config.hidden_size, self.padding_idx)
+        self.layers = nn.ModuleList([RWKV7Block(config, layer_idx) for layer_idx in range(config.num_hidden_layers)])
+        self.norm = (LayerNorm if config.fuse_norm else nn.LayerNorm)(
+            config.hidden_size,
+            bias=config.norm_bias,
+            eps=config.norm_eps
+        )
+
+        self.gradient_checkpointing = False
+
+        self.post_init()
+
+    def get_input_embeddings(self):
+        return self.embeddings
+
+    def set_input_embeddings(self, value):
+        self.embeddings = value
+
+    def forward(
+        self,
+        input_ids: Optional[torch.LongTensor] = None,
+        attention_mask: Optional[torch.Tensor] = None,  # noqa
+        inputs_embeds: Optional[torch.FloatTensor] = None,
+        past_key_values: Optional[Cache] = None,
+        use_cache: Optional[bool] = None,
+        output_attentions: Optional[bool] = None,
+        output_hidden_states: Optional[bool] = None,
+        return_dict: Optional[bool] = None,
+        **kwargs: Unpack[Dict]
+    ) -> Union[Tuple, BaseModelOutputWithPast]:
+        if output_attentions:
+            warnings.warn("`RWKV7Model` does not `output_attentions` now, setting it to `False`.")
+            output_attentions = False
+        output_attentions = output_attentions if output_attentions is not None else self.config.output_attentions
+        output_hidden_states = output_hidden_states if output_hidden_states is not None else self.config.output_hidden_states
+        use_cache = use_cache if use_cache is not None else (self.config.use_cache if not self.training else False)
+        return_dict = return_dict if return_dict is not None else self.config.use_return_dict
+
+        # retrieve input_ids and inputs_embeds
+        if input_ids is not None and inputs_embeds is not None:
+            raise ValueError("You cannot specify both input_ids and inputs_embeds at the same time")
+        if input_ids is None and inputs_embeds is None:
+            raise ValueError("You have to specify either input_ids or inputs_embeds")
+
+        if inputs_embeds is None:
+            inputs_embeds = self.embeddings(input_ids)
+        hidden_states = inputs_embeds
+
+        if use_cache and not isinstance(past_key_values, Cache):
+            past_key_values = Cache.from_legacy_cache(past_key_values)
+
+        if self.gradient_checkpointing and self.training and use_cache:
+            logger.warning_once("`use_cache=True` is incompatible with gradient checkpointing. Setting `use_cache=False`...")
+            use_cache = False
+
+        all_hidden_states = () if output_hidden_states else None
+        all_attns = () if output_attentions else None
+
+        v_first = torch.zeros_like(hidden_states)
+        for layer in self.layers:
+            if output_hidden_states:
+                all_hidden_states += (hidden_states,)
+
+            if self.gradient_checkpointing and self.training:
+                hidden_states, attentions, past_key_values, v_first = self._gradient_checkpointing_func(
+                    layer.__call__,
+                    hidden_states,
+                    attention_mask,
+                    past_key_values,
+                    use_cache,
+                    output_attentions,
+                    v_first,
+                    **kwargs
+                )
+            else:
+                hidden_states, attentions, past_key_values, v_first = layer(
+                    hidden_states,
+                    attention_mask=attention_mask,
+                    past_key_values=past_key_values,
+                    use_cache=use_cache,
+                    output_attentions=output_attentions,
+                    v_first=v_first,
+                    **kwargs
+                )
+
+            if output_attentions:
+                all_attns += (attentions,)
+
+        hidden_states = self.norm(hidden_states)
+
+        # add hidden states from the last decoder layer
+        if output_hidden_states:
+            all_hidden_states += (hidden_states,)
+
+        if not return_dict:
+            return tuple(i for i in [hidden_states, past_key_values, all_hidden_states, all_attns] if i is not None)
+        return BaseModelOutputWithPast(
+            last_hidden_state=hidden_states,
+            past_key_values=past_key_values,
+            hidden_states=all_hidden_states,
+            attentions=all_attns
+        )
+
+
+class RWKV7ForCausalLM(RWKV7PreTrainedModel, GenerationMixin):
+
+    _tied_weights_keys = ["lm_head.weight"]
+
+    def __init__(self, config):
+        super().__init__(config)
+        self.model = RWKV7Model(config)
+        self.vocab_size = config.vocab_size
+        self.lm_head = nn.Linear(config.hidden_size, config.vocab_size, bias=False)
+        self.criterion = None
+
+        # Initialize weights and apply final processing
+        self.post_init()
+
+    def get_input_embeddings(self):
+        return self.model.embeddings
+
+    def set_input_embeddings(self, value):
+        self.model.embeddings = value
+
+    def get_output_embeddings(self):
+        return self.lm_head
+
+    def set_output_embeddings(self, new_embeddings):
+        self.lm_head = new_embeddings
+
+    def set_decoder(self, decoder):
+        self.model = decoder
+
+    def get_decoder(self):
+        return self.model
+
+    def generate(self, *args, **kwargs):
+        try:
+            return super().generate(*args, **kwargs)
+        except AttributeError as exception:
+            if 'past_key_values' in str(exception):
+                raise AttributeError(
+                    f"You tried to call `generate` with a decoding strategy that manipulates `past_key_values`, "
+                    f"which is not supported for {self.__class__.__name__}. "
+                    f"Try another generation strategy instead. "
+                    f"For the available generation strategies, check this doc: "
+                    f"https://huggingface.co/docs/transformers/en/generation_strategies#decoding-strategies"
+                )
+            else:
+                raise exception
+
+    @deprecate_kwarg("num_logits_to_keep", version="4.50", new_name="logits_to_keep")
+    def prepare_inputs_for_generation(
+        self,
+        input_ids: torch.LongTensor = None,
+        past_key_values: Optional[Cache] = None,
+        attention_mask: Optional[torch.Tensor] = None,
+        inputs_embeds: Optional[torch.Tensor] = None,
+        use_cache: bool = True,
+        logits_to_keep: Optional[int] = None,
+        **kwargs
+    ):
+        # only last token for `inputs_ids` if the `past_key_values` is not empty.
+        if past_key_values is not None and len(past_key_values) > 0:
+            input_ids = input_ids[:, -1:]
+        # if `inputs_embeds` are passed, we only want to use them in the 1st generation step
+        if inputs_embeds is not None and len(past_key_values) == 0:
+            model_inputs = {'inputs_embeds': inputs_embeds}
+        else:
+            # The `contiguous()` here is necessary to have a static stride during decoding. torchdynamo otherwise
+            # recompiles graphs as the stride of the inputs is a guard.
+            # Ref: https://github.com/huggingface/transformers/pull/29114
+            # TODO: use `next_tokens` directly instead.
+            model_inputs = {'input_ids': input_ids.contiguous()}
+
+        if logits_to_keep is not None:
+            model_inputs['logits_to_keep'] = logits_to_keep
+
+        model_inputs.update({
+            'past_key_values': past_key_values,
+            'use_cache': use_cache,
+            'attention_mask': attention_mask,
+        })
+        return model_inputs
+
+    @deprecate_kwarg("num_logits_to_keep", version="4.50", new_name="logits_to_keep")
+    def forward(
+        self,
+        input_ids: torch.LongTensor = None,
+        attention_mask: Optional[torch.Tensor] = None,
+        inputs_embeds: Optional[torch.Tensor] = None,
+        past_key_values: Optional[Cache] = None,
+        labels: Optional[torch.LongTensor] = None,
+        shift_labels: Optional[torch.LongTensor] = None,
+        use_cache: Optional[bool] = None,
+        output_attentions: Optional[bool] = None,
+        output_hidden_states: Optional[bool] = None,
+        return_dict: Optional[bool] = None,
+        logits_to_keep: Optional[int] = 0,
+        **kwargs: Unpack[Dict]
+    ) -> Union[Tuple, CausalLMOutputWithPast]:
+        output_attentions = output_attentions if output_attentions is not None else self.config.output_attentions
+        output_hidden_states = (
+            output_hidden_states if output_hidden_states is not None else self.config.output_hidden_states
+        )
+        return_dict = return_dict if return_dict is not None else self.config.use_return_dict
+
+        outputs = self.model(
+            input_ids=input_ids,
+            attention_mask=attention_mask,
+            inputs_embeds=inputs_embeds,
+            past_key_values=past_key_values,
+            use_cache=use_cache,
+            output_attentions=output_attentions,
+            output_hidden_states=output_hidden_states,
+            return_dict=return_dict,
+            **kwargs
+        )
+
+        hidden_states = outputs[0]
+        fuse_linear_and_cross_entropy = self.config.fuse_cross_entropy and self.training
+
+        loss, logits = None, None
+        has_labels = (labels is not None) or (shift_labels is not None)
+        if not (fuse_linear_and_cross_entropy and has_labels):
+            logits = self.lm_head(hidden_states if logits_to_keep is None else hidden_states[:, -logits_to_keep:])
+        if has_labels:
+            if getattr(self, 'criterion', None) is None:
+                if fuse_linear_and_cross_entropy:
+                    criterion = FusedLinearCrossEntropyLoss()
+                elif self.config.fuse_cross_entropy:
+                    criterion = FusedCrossEntropyLoss(inplace_backward=True)
+                else:
+                    criterion = nn.CrossEntropyLoss()
+            else:
+                criterion = self.criterion
+
+            # shift_labels: See https://github.com/huggingface/transformers/pull/36607/files.
+            if shift_labels is None:
+                shift_labels = torch.cat((labels[..., 1:], torch.full_like(labels[:, :1], criterion.ignore_index)), 1)
+            shift_labels = shift_labels.to(hidden_states.device)
+
+            if fuse_linear_and_cross_entropy:
+                loss = criterion(hidden_states, shift_labels, self.lm_head.weight, self.lm_head.bias)
+            else:
+                loss = criterion(logits.view(shift_labels.numel(), -1), shift_labels.view(-1))
+
+        if not return_dict:
+            output = (logits,) + outputs[1:]
+            return (loss,) + output if loss is not None else output
+
+        return CausalLMOutputWithPast(
+            loss=loss,
+            logits=logits,
+            past_key_values=outputs.past_key_values,
+            hidden_states=outputs.hidden_states,
+            attentions=outputs.attentions,
+        )

fla/models/samba/__init__.py

@@ -0,0 +1,13 @@
+# -*- coding: utf-8 -*-
+
+from transformers import AutoConfig, AutoModel, AutoModelForCausalLM
+
+from fla.models.samba.configuration_samba import SambaConfig
+from fla.models.samba.modeling_samba import SambaBlock, SambaForCausalLM, SambaModel
+
+AutoConfig.register(SambaConfig.model_type, SambaConfig, True)
+AutoModel.register(SambaConfig, SambaModel, True)
+AutoModelForCausalLM.register(SambaConfig, SambaForCausalLM, True)
+
+
+__all__ = ['SambaConfig', 'SambaForCausalLM', 'SambaModel', 'SambaBlock']

fla/models/samba/configuration_samba.py

@@ -0,0 +1,92 @@
+# -*- coding: utf-8 -*-
+
+import math
+from typing import Dict, Optional
+
+from transformers.configuration_utils import PretrainedConfig
+
+
+class SambaConfig(PretrainedConfig):
+
+    model_type = "samba"
+
+    def __init__(
+        self,
+        hidden_size: int = 2304,
+        state_size: int = 16,
+        num_hidden_layers: int = 18,
+        norm_eps=1e-5,
+        pad_token_id: int = 0,
+        bos_token_id: int = 1,
+        eos_token_id: int = 2,
+        expand: int = 2,
+        conv_kernel: int = 4,
+        use_bias: bool = False,
+        use_conv_bias: bool = True,
+        hidden_act: str = "swish",
+        initializer_range: str = 0.02,
+        residual_in_fp32: bool = False,
+        time_step_rank: str = "auto",
+        time_step_scale: float = 1.0,
+        time_step_min: float = 0.001,
+        time_step_max: float = 0.1,
+        time_step_init_scheme: str = "random",
+        time_step_floor: float = 1e-4,
+        max_position_embeddings: int = 2048,
+        attn: Optional[Dict] = {
+            'layers': (1, 3, 5, 7, 9, 11, 13, 15, 17),
+            'num_heads': 18,
+            'num_kv_heads': 18,
+            'qkv_bias': False,
+            'window_size': 2048,
+            'rope_theta': 10000.
+        },
+        hidden_ratio: Optional[int] = 4,
+        rescale_prenorm_residual: bool = False,
+        use_cache: bool = True,
+        fuse_norm: bool = True,
+        fuse_swiglu: bool = True,
+        fuse_cross_entropy: bool = True,
+        vocab_size: int = 32000,
+        tie_word_embeddings: bool = False,
+        **kwargs,
+    ):
+        self.hidden_size = hidden_size
+        self.state_size = state_size
+        self.num_hidden_layers = num_hidden_layers
+        self.norm_eps = norm_eps
+        self.conv_kernel = conv_kernel
+        self.expand = expand
+        self.intermediate_size = int(expand * self.hidden_size)
+        self.bos_token_id = bos_token_id
+        self.eos_token_id = eos_token_id
+        self.pad_token_id = pad_token_id
+        self.use_bias = use_bias
+        self.use_conv_bias = use_conv_bias
+        self.hidden_act = hidden_act
+        self.initializer_range = initializer_range
+        self.time_step_rank = math.ceil(self.hidden_size / 16) if time_step_rank == "auto" else time_step_rank
+        self.time_step_scale = time_step_scale
+        self.time_step_min = time_step_min
+        self.time_step_max = time_step_max
+        self.time_step_init_scheme = time_step_init_scheme
+        self.time_step_floor = time_step_floor
+        self.max_position_embeddings = max_position_embeddings
+        self.attn = attn
+        self.hidden_ratio = hidden_ratio
+        self.rescale_prenorm_residual = rescale_prenorm_residual
+        self.residual_in_fp32 = residual_in_fp32
+        self.use_cache = use_cache
+
+        self.fuse_norm = fuse_norm
+        self.fuse_swiglu = fuse_swiglu
+        self.fuse_cross_entropy = fuse_cross_entropy
+        self.vocab_size = vocab_size
+
+        super().__init__(
+            bos_token_id=bos_token_id,
+            eos_token_id=eos_token_id,
+            pad_token_id=pad_token_id,
+            tie_word_embeddings=tie_word_embeddings,
+            **kwargs
+        )

fla/models/samba/modeling_samba.py

@@ -0,0 +1,413 @@
+# -*- coding: utf-8 -*-
+
+from __future__ import annotations
+
+import math
+from dataclasses import dataclass
+from typing import TYPE_CHECKING, Any, Dict, Optional, Tuple, Union
+
+import torch
+import torch.utils.checkpoint
+from torch import nn
+from transformers.generation import GenerationMixin
+from transformers.modeling_utils import PreTrainedModel
+from transformers.utils import ModelOutput, logging
+from transformers.utils.deprecation import deprecate_kwarg
+
+from fla.layers.attn import Attention
+from fla.models.mamba.modeling_mamba import MambaCache, MambaMixer
+from fla.models.samba.configuration_samba import SambaConfig
+from fla.modules import FusedCrossEntropyLoss, FusedLinearCrossEntropyLoss
+from fla.modules import GatedMLP as SambaMLP
+from fla.modules import RMSNorm
+
+if TYPE_CHECKING:
+    from transformers.processing_utils import Unpack
+
+logger = logging.get_logger(__name__)
+
+
+class SambaBlock(nn.Module):
+    def __init__(self, config, layer_idx):
+        super().__init__()
+
+        self.config = config
+        self.layer_idx = layer_idx
+
+        self.mixer_norm = RMSNorm(hidden_size=config.hidden_size, eps=config.norm_eps)
+        if config.attn is not None and layer_idx in config.attn['layers']:
+            self.mixer = Attention(
+                hidden_size=config.hidden_size,
+                num_heads=config.attn['num_heads'],
+                num_kv_heads=config.attn['num_kv_heads'],
+                qkv_bias=config.attn['qkv_bias'],
+                window_size=config.attn['window_size'],
+                rope_theta=config.attn['rope_theta'],
+                max_position_embeddings=config.max_position_embeddings,
+                layer_idx=layer_idx
+            )
+        else:
+            self.mixer = MambaMixer(config, layer_idx=layer_idx)
+        self.mlp_norm = (RMSNorm if config.fuse_norm else nn.RMSNorm)(config.hidden_size, eps=config.norm_eps)
+        self.mlp = SambaMLP(
+            hidden_size=config.hidden_size,
+            hidden_ratio=config.hidden_ratio,
+            hidden_act=config.hidden_act,
+            fuse_swiglu=config.fuse_swiglu
+        )
+
+    def forward(
+        self,
+        hidden_states: torch.Tensor,
+        cache_params: Optional[Tuple[torch.Tensor]] = None,
+        **kwargs: Unpack[Dict]
+    ) -> Tuple[torch.FloatTensor, Optional[Tuple[torch.FloatTensor, torch.FloatTensor]]]:
+
+        residual = hidden_states
+        hidden_states = self.mixer_norm(hidden_states)
+        if isinstance(self.mixer, MambaMixer):
+            hidden_states = self.mixer(hidden_states, cache_params=cache_params, **kwargs)
+        else:
+            hidden_states, _, cache_params = self.mixer(hidden_states=hidden_states, past_key_values=cache_params, **kwargs)
+        if self.config.fuse_norm:
+            hidden_states, residual = self.mlp_norm(hidden_states, residual, True)
+        else:
+            hidden_states = residual + hidden_states
+            residual = hidden_states
+            hidden_states = self.mlp_norm(hidden_states)
+        hidden_states = self.mlp(hidden_states, **kwargs)
+        hidden_states = residual + hidden_states
+        return hidden_states
+
+
+class SambaPreTrainedModel(PreTrainedModel):
+    """
+    An abstract class to handle weights initialization and a simple interface for downloading and loading pretrained
+    models.
+    """
+
+    config_class = SambaConfig
+    base_model_prefix = "backbone"
+    _no_split_modules = ["SambaBlock"]
+    supports_gradient_checkpointing = True
+
+    def _init_weights(self, module):
+        """Initialize the weights."""
+        if isinstance(module, nn.Linear):
+            nn.init.normal_(module.weight, mean=0.0, std=self.config.initializer_range)
+            if module.bias is not None:
+                if not getattr(module.bias, "_no_reinit", False):
+                    nn.init.zeros_(module.bias)
+        elif isinstance(module, MambaMixer):
+            module.A_log._no_weight_decay = True
+            module.D._no_weight_decay = True
+
+            dt_init_std = self.config.time_step_rank**-0.5 * self.config.time_step_scale
+            if self.config.time_step_init_scheme == "constant":
+                nn.init.constant_(module.dt_proj.weight, dt_init_std)
+            elif self.config.time_step_init_scheme == "random":
+                nn.init.uniform_(module.dt_proj.weight, -dt_init_std, dt_init_std)
+
+            dt = torch.exp(
+                torch.rand(self.config.intermediate_size)
+                * (math.log(self.config.time_step_max) - math.log(self.config.time_step_min))
+                + math.log(self.config.time_step_min)
+            ).clamp(min=self.config.time_step_floor)
+            # # Inverse of softplus: https://github.com/pytorch/pytorch/issues/72759
+            inv_dt = dt + torch.log(-torch.expm1(-dt))
+            with torch.no_grad():
+                module.dt_proj.bias.data = nn.Parameter(inv_dt.to(module.dt_proj.bias.device))
+            module.dt_proj.bias._no_reinit = True
+        elif isinstance(module, nn.Embedding):
+            nn.init.normal_(module.weight, std=self.config.initializer_range)
+        elif hasattr(module, 'reset_parameters'):
+            module.reset_parameters()
+
+        if self.config.rescale_prenorm_residual:
+            # Reinitialize selected weights subject to the OpenAI GPT-2 Paper Scheme:
+            #   > A modified initialization which accounts for the accumulation on the residual path with model depth. Scale
+            #   > the weights of residual layers at initialization by a factor of 1/√N where N is the # of residual layers.
+            #   >   -- GPT-2 :: https://openai.com/blog/better-language-models/
+            #
+            # Reference (Megatron-LM): https://github.com/NVIDIA/Megatron-LM/blob/main/megatron/model/gpt_model.py
+            for name, p in module.named_parameters():
+                if name in ["out_proj.weight"]:
+                    # Special Scaled Initialization --> There are 2 Layer Norms per Transformer Block
+                    # Following Pytorch init, except scale by 1/sqrt(2 * n_layer)
+                    # We need to reinit p since this code could be called multiple times
+                    # Having just p *= scale would repeatedly scale it down
+                    nn.init.kaiming_uniform_(p, a=math.sqrt(5))
+                    with torch.no_grad():
+                        p /= math.sqrt(self.config.num_layers)
+
+
+@dataclass
+class SambaOutput(ModelOutput):
+    """
+    Class for the Samba model outputs.
+
+    Args:
+        last_hidden_state (`torch.FloatTensor` of shape `(batch_size, sequence_length, hidden_size)`):
+            Sequence of hidden-states at the output of the last layer of the model.
+        cache_params (`MambaCache`):
+            The state of the model at the last time step. Can be used in a forward method with the next `input_ids` to
+            avoid providing the old `input_ids`.
+
+            Includes both the State space model state matrices after the selective scan, and the Convolutional states
+        hidden_states (`tuple(torch.FloatTensor)`, *optional*,
+            returned when `output_hidden_states=True` is passed or when `config.output_hidden_states=True`):
+            Tuple of `torch.FloatTensor` (one for the output of the embeddings, if the model has an embedding layer, +
+            one for the output of each layer) of shape `(batch_size, sequence_length, hidden_size)`.
+
+            Hidden-states of the model at the output of each layer plus the optional initial embedding outputs.
+    """
+
+    last_hidden_state: Optional[torch.FloatTensor] = None
+    cache_params: Optional[MambaCache] = None
+    hidden_states: Optional[Tuple[torch.FloatTensor]] = None
+
+
+@dataclass
+class SambaCausalLMOutput(ModelOutput):
+    """
+    Base class for causal language model (or autoregressive) outputs.
+
+    Args:
+        loss (`torch.FloatTensor` of shape `(1,)`, *optional*, returned when `labels` is provided):
+            Language modeling loss (for next-token prediction).
+        logits (`torch.FloatTensor` of shape `(batch_size, sequence_length, config.vocab_size)`):
+            Prediction scores of the language modeling head (scores for each vocabulary token before SoftMax).
+        cache_params (`MambaCache`):
+            The state of the model at the last time step. Can be used in a forward method with the next `input_ids` to
+            avoid providing the old `input_ids`.
+
+            Includes both the State space model state matrices after the selective scan, and the Convolutional states
+        hidden_states (`tuple(torch.FloatTensor)`, *optional*,
+            returned when `output_hidden_states=True` is passed or when `config.output_hidden_states=True`):
+            Tuple of `torch.FloatTensor` (one for the output of the embeddings, if the model has an embedding layer, +
+            one for the output of each layer) of shape `(batch_size, sequence_length, hidden_size)`.
+
+            Hidden-states of the model at the output of each layer plus the optional initial embedding outputs.
+    """
+
+    loss: Optional[torch.FloatTensor] = None
+    logits: Optional[torch.FloatTensor] = None
+    cache_params: Optional[MambaCache] = None
+    hidden_states: Optional[Tuple[torch.FloatTensor]] = None
+
+
+class SambaModel(SambaPreTrainedModel):
+    def __init__(self, config):
+        super().__init__(config)
+
+        self.embeddings = nn.Embedding(config.vocab_size, config.hidden_size)
+        self.layers = nn.ModuleList([SambaBlock(config, layer_idx=idx) for idx in range(config.num_hidden_layers)])
+
+        self.gradient_checkpointing = False
+        self.norm_f = RMSNorm(config.hidden_size, eps=config.norm_eps)
+        # Initialize weights and apply final processing
+        self.post_init()
+
+    def get_input_embeddings(self):
+        return self.embeddings
+
+    def set_input_embeddings(self, new_embeddings):
+        self.embeddings = new_embeddings
+
+    def forward(
+        self,
+        input_ids: Optional[torch.LongTensor] = None,
+        inputs_embeds: Optional[torch.LongTensor] = None,
+        cache_params: Optional[MambaCache] = None,
+        use_cache: Optional[bool] = None,
+        output_hidden_states: Optional[bool] = None,
+        return_dict: Optional[bool] = None,
+        **kwargs: Unpack[Dict]
+    ) -> Union[Tuple, SambaOutput]:
+        output_hidden_states = (
+            output_hidden_states if output_hidden_states is not None else self.config.output_hidden_states
+        )
+        use_cache = use_cache if use_cache is not None else (self.config.use_cache if not self.training else False)
+        return_dict = return_dict if return_dict is not None else self.config.use_return_dict
+
+        if (input_ids is None) ^ (inputs_embeds is not None):  # ^ is python for xor
+            raise ValueError(
+                "You cannot specify both input_ids and inputs_embeds at the same time, and must specify either one"
+            )
+
+        if inputs_embeds is None:
+            inputs_embeds = self.embeddings(input_ids)
+
+        if self.gradient_checkpointing and self.training and use_cache:
+            use_cache = False
+
+        if cache_params is None and use_cache:
+            cache_params = MambaCache(
+                self.config, inputs_embeds.size(0), device=inputs_embeds.device, dtype=inputs_embeds.dtype
+            )
+
+        hidden_states = inputs_embeds
+        all_hidden_states = () if output_hidden_states else None
+        for mixer_block in self.layers:
+            if self.gradient_checkpointing and self.training:
+                hidden_states = self._gradient_checkpointing_func(
+                    mixer_block.__call__,
+                    hidden_states,
+                    cache_params,
+                    **kwargs
+                )
+            else:
+                hidden_states = mixer_block(
+                    hidden_states,
+                    cache_params=cache_params,
+                    **kwargs
+                )
+
+            if output_hidden_states:
+                all_hidden_states = all_hidden_states + (hidden_states,)
+
+        if use_cache:
+            cache_params.seqlen_offset += inputs_embeds.shape[1]
+
+        hidden_states = self.norm_f(hidden_states)
+
+        if output_hidden_states:
+            all_hidden_states = all_hidden_states + (hidden_states,)
+
+        if not return_dict:
+            return tuple(v for v in [hidden_states, cache_params, all_hidden_states] if v is not None)
+
+        return SambaOutput(
+            last_hidden_state=hidden_states,
+            cache_params=cache_params if use_cache else None,
+            hidden_states=all_hidden_states,
+        )
+
+
+class SambaForCausalLM(SambaPreTrainedModel, GenerationMixin):
+
+    _tied_weights_keys = ["lm_head.weight"]
+
+    def __init__(self, config):
+        super().__init__(config)
+        self.backbone = SambaModel(config)
+        self.lm_head = nn.Linear(config.hidden_size, config.vocab_size, bias=False)
+        self.criterion = None
+
+        # Initialize weights and apply final processing
+        self.post_init()
+
+    def get_output_embeddings(self):
+        return self.lm_head
+
+    def set_output_embeddings(self, new_embeddings):
+        self.lm_head = new_embeddings
+
+    def get_input_embeddings(self):
+        return self.backbone.get_input_embeddings()
+
+    def set_input_embeddings(self, new_embeddings):
+        return self.backbone.set_input_embeddings(new_embeddings)
+
+    def _update_model_kwargs_for_generation(
+        self, outputs: ModelOutput, model_kwargs: Dict[str, Any], **kwargs
+    ) -> Dict[str, Any]:
+        model_kwargs["cache_params"] = outputs.get("cache_params", None)
+        return model_kwargs
+
+    @deprecate_kwarg("num_logits_to_keep", version="4.50", new_name="logits_to_keep")
+    def prepare_inputs_for_generation(
+        self,
+        input_ids,
+        cache_params:
+        Optional[MambaCache] = None,
+        inputs_embeds=None,
+        attention_mask=None,
+        use_cache: Optional[bool] = True,
+        logits_to_keep: Optional[int] = None,
+        **kwargs: Unpack[Dict]
+    ):
+        # only last token for inputs_ids if the state is passed along.
+        if cache_params is not None:
+            input_ids = input_ids[:, -1].unsqueeze(-1)
+
+        if inputs_embeds is not None and cache_params is None:
+            model_inputs = {"inputs_embeds": inputs_embeds}
+        else:
+            model_inputs = {"input_ids": input_ids}
+
+        if logits_to_keep is not None:
+            model_inputs['logits_to_keep'] = logits_to_keep
+
+        model_inputs.update({
+            'cache_params': cache_params,
+            'use_cache': use_cache,
+            'attention_mask': attention_mask,
+            'logits_to_keep': logits_to_keep,
+        })
+        return model_inputs
+
+    @deprecate_kwarg("num_logits_to_keep", version="4.50", new_name="logits_to_keep")
+    def forward(
+        self,
+        input_ids: Optional[torch.LongTensor] = None,
+        attention_mask: Optional[torch.Tensor] = None,  # noqa
+        inputs_embeds: Optional[torch.FloatTensor] = None,
+        cache_params: Optional[MambaCache] = None,
+        labels: Optional[torch.LongTensor] = None,
+        output_hidden_states: Optional[bool] = None,
+        return_dict: Optional[bool] = None,
+        use_cache: Optional[bool] = None,
+        logits_to_keep: Optional[int] = 0,
+        **kwargs: Unpack[Dict]
+    ) -> Union[Tuple, SambaCausalLMOutput]:
+        r"""
+        labels (`torch.LongTensor` of shape `(batch_size, sequence_length)`, *optional*):
+            Labels for language modeling. Note that the labels **are shifted** inside the model, i.e. you can set
+            `labels = input_ids` Indices are selected in `[-100, 0, ..., config.vocab_size]` All labels set to `-100`
+            are ignored (masked), the loss is only computed for labels in `[0, ..., config.vocab_size]`
+        """
+        return_dict = return_dict if return_dict is not None else self.config.use_return_dict
+
+        outputs = self.backbone(
+            input_ids,
+            cache_params=cache_params,
+            inputs_embeds=inputs_embeds,
+            output_hidden_states=output_hidden_states,
+            return_dict=return_dict,
+            use_cache=use_cache,
+            **kwargs
+        )
+        hidden_states = outputs[0]
+        fuse_linear_and_cross_entropy = self.config.fuse_cross_entropy and self.training
+
+        loss, logits = None, None
+        if not fuse_linear_and_cross_entropy or labels is None:
+            logits = self.lm_head(hidden_states if logits_to_keep is None else hidden_states[:, -logits_to_keep:])
+        if labels is not None:
+            if getattr(self, 'criterion', None) is None:
+                if fuse_linear_and_cross_entropy:
+                    criterion = FusedLinearCrossEntropyLoss()
+                elif self.config.fuse_cross_entropy:
+                    criterion = FusedCrossEntropyLoss(inplace_backward=True)
+                else:
+                    criterion = nn.CrossEntropyLoss()
+            else:
+                criterion = self.criterion
+            labels = labels.to(hidden_states.device)
+            labels = torch.cat((labels[..., 1:], torch.full_like(labels[:, :1], criterion.ignore_index)), 1)
+            if fuse_linear_and_cross_entropy:
+                loss = criterion(hidden_states, labels, self.lm_head.weight, self.lm_head.bias)
+            else:
+                loss = criterion(logits.view(labels.numel(), -1), labels.view(-1))
+
+        if not return_dict:
+            output = (logits,) + outputs[1:]
+            return (loss,) + output if loss is not None else output
+
+        return SambaCausalLMOutput(
+            loss=loss,
+            logits=logits,
+            cache_params=outputs.cache_params,
+            hidden_states=outputs.hidden_states,
+        )

fla/models/transformer/modeling_transformer.py

@@ -0,0 +1,406 @@
+# -*- coding: utf-8 -*-
+
+from __future__ import annotations
+
+import math
+import warnings
+from typing import TYPE_CHECKING, Any, List, Optional, Tuple, Union
+
+import torch
+import torch.nn as nn
+import torch.utils.checkpoint
+from transformers.generation import GenerationMixin
+from transformers.modeling_outputs import BaseModelOutputWithPast, CausalLMOutputWithPast
+from transformers.modeling_utils import PreTrainedModel
+from transformers.utils import logging
+from transformers.utils.deprecation import deprecate_kwarg
+
+from fla.layers.attn import Attention
+from fla.models.transformer.configuration_transformer import TransformerConfig
+from fla.models.utils import Cache
+from fla.modules import FusedCrossEntropyLoss, FusedLinearCrossEntropyLoss
+from fla.modules import GatedMLP as TransformerMLP
+from fla.modules import RMSNorm
+
+if TYPE_CHECKING:
+    from transformers.processing_utils import Unpack
+
+
+logger = logging.get_logger(__name__)
+
+
+class TransformerBlock(nn.Module):
+
+    def __init__(self, config: TransformerConfig, layer_idx: int):
+        super().__init__()
+
+        self.config = config
+        self.layer_idx = layer_idx
+
+        self.attn_norm = (RMSNorm if config.fuse_norm else nn.RMSNorm)(config.hidden_size, eps=config.norm_eps)
+        self.attn = Attention(
+            hidden_size=config.hidden_size,
+            num_heads=config.num_heads,
+            num_kv_heads=config.num_kv_heads,
+            qkv_bias=config.qkv_bias,
+            qk_norm=config.qk_norm,
+            window_size=config.window_size,
+            rope_theta=config.rope_theta,
+            max_position_embeddings=config.max_position_embeddings,
+            layer_idx=layer_idx
+        )
+
+        self.mlp_norm = (RMSNorm if config.fuse_norm else nn.RMSNorm)(config.hidden_size, eps=config.norm_eps)
+        self.mlp = TransformerMLP(
+            hidden_size=config.hidden_size,
+            hidden_ratio=config.hidden_ratio,
+            intermediate_size=config.intermediate_size,
+            hidden_act=config.hidden_act,
+            fuse_swiglu=config.fuse_swiglu
+        )
+
+    def forward(
+        self,
+        hidden_states: torch.Tensor,
+        attention_mask: Optional[torch.Tensor] = None,
+        past_key_values: Optional[Tuple[torch.Tensor]] = None,
+        output_attentions: Optional[bool] = False,
+        use_cache: Optional[bool] = False,
+        **kwargs: Unpack[Any]
+    ) -> Tuple[torch.FloatTensor, Optional[Tuple[torch.FloatTensor, torch.FloatTensor]]]:
+
+        residual = hidden_states
+        hidden_states = self.attn_norm(hidden_states)
+        hidden_states, attentions, past_key_values = self.attn(
+            hidden_states=hidden_states,
+            attention_mask=attention_mask,
+            past_key_values=past_key_values,
+            use_cache=use_cache,
+            output_attentions=output_attentions,
+            **kwargs
+        )
+        if self.config.fuse_norm:
+            hidden_states, residual = self.mlp_norm(hidden_states, residual, True)
+        else:
+            hidden_states = residual + hidden_states
+            residual = hidden_states
+            hidden_states = self.mlp_norm(hidden_states)
+        hidden_states = self.mlp(hidden_states, **kwargs)
+        hidden_states = residual + hidden_states
+
+        outputs = (hidden_states,)
+
+        if output_attentions:
+            outputs += (attentions,)
+
+        if use_cache:
+            outputs += (past_key_values,)
+
+        return outputs
+
+
+class TransformerPreTrainedModel(PreTrainedModel):
+
+    config_class = TransformerConfig
+    base_model_prefix = 'model'
+    supports_gradient_checkpointing = True
+    _no_split_modules = ['TransformerBlock']
+    _supports_cache_class = True
+
+    def __init__(self, *inputs, **kwargs):
+        super().__init__(*inputs, **kwargs)
+
+    def _init_weights(
+        self,
+        module: nn.Module,
+        rescale_prenorm_residual: bool = False,
+        num_residuals_per_layer: int = 2,
+    ):
+        if isinstance(module, (nn.Linear, nn.Conv1d)):
+            # Slightly different from the TF version which uses truncated_normal for initialization
+            # cf https://github.com/pytorch/pytorch/pull/5617
+            nn.init.normal_(module.weight, mean=0.0, std=self.config.initializer_range)
+            if module.bias is not None:
+                nn.init.zeros_(module.bias)
+        elif isinstance(module, nn.Embedding):
+            nn.init.normal_(module.weight, mean=0.0, std=self.config.initializer_range)
+        elif hasattr(module, 'reset_parameters'):
+            module.reset_parameters()
+
+        if rescale_prenorm_residual:
+            # Reinitialize selected weights subject to the OpenAI GPT-2 Paper Scheme:
+            #   > A modified initialization which accounts for the accumulation on the residual path with model depth. Scale
+            #   > the weights of residual layers at initialization by a factor of 1/√N where N is the # of residual layers.
+            #   >   -- GPT-2 :: https://openai.com/blog/better-language-models/
+            #
+            # Reference (Megatron-LM): https://github.com/NVIDIA/Megatron-LM/blob/main/megatron/model/gpt_model.py
+            p = None
+            if hasattr(module, 'o_proj'):
+                p = module.o_proj.weight
+            elif hasattr(module, 'down_proj'):
+                p = module.down_proj.weight
+            if p is not None:
+                # Special Scaled Initialization --> There are 2 Layer Norms per Transformer Block
+                # Following Pytorch init, except scale by 1/sqrt(2 * n_layer)
+                # We need to reinit p since this code could be called multiple times
+                # Having just p *= scale would repeatedly scale it down
+                nn.init.kaiming_uniform_(p, a=math.sqrt(5))
+                with torch.no_grad():
+                    p /= math.sqrt(num_residuals_per_layer * self.config.num_hidden_layers)
+
+
+class TransformerModel(TransformerPreTrainedModel):
+
+    def __init__(
+        self,
+        config: TransformerConfig
+    ) -> TransformerModel:
+        super().__init__(config)
+        self.padding_idx = config.pad_token_id
+        self.vocab_size = config.vocab_size
+
+        self.embeddings = nn.Embedding(config.vocab_size, config.hidden_size, self.padding_idx)
+        self.layers = nn.ModuleList([TransformerBlock(config, layer_idx) for layer_idx in range(config.num_hidden_layers)])
+        self.norm = (RMSNorm if config.fuse_norm else nn.RMSNorm)(config.hidden_size, eps=config.norm_eps)
+
+        self.gradient_checkpointing = False
+
+        self.post_init()
+
+    def get_input_embeddings(self):
+        return self.embeddings
+
+    def set_input_embeddings(self, value):
+        self.embeddings = value
+
+    def forward(
+        self,
+        input_ids: Optional[torch.LongTensor] = None,
+        attention_mask: Optional[torch.Tensor] = None,
+        past_key_values: Optional[List[torch.FloatTensor]] = None,
+        inputs_embeds: Optional[torch.FloatTensor] = None,
+        use_cache: Optional[bool] = None,
+        output_attentions: Optional[bool] = None,
+        output_hidden_states: Optional[bool] = None,
+        return_dict: Optional[bool] = None,
+        **kwargs: Unpack[Any]
+    ) -> Union[Tuple, CausalLMOutputWithPast]:
+        if output_attentions:
+            warnings.warn(
+                "`TransformerModel` does not support output attention weights now, so `output_attentions` is set to `False`."
+            )
+            output_attentions = False
+        output_attentions = output_attentions if output_attentions is not None else self.config.output_attentions
+        output_hidden_states = output_hidden_states if output_hidden_states is not None else self.config.output_hidden_states
+        use_cache = use_cache if use_cache is not None else (self.config.use_cache if not self.training else False)
+        return_dict = return_dict if return_dict is not None else self.config.use_return_dict
+
+        # retrieve input_ids and inputs_embeds
+        if input_ids is not None and inputs_embeds is not None:
+            raise ValueError("You cannot specify both input_ids and inputs_embeds at the same time")
+        elif input_ids is None and inputs_embeds is None:
+            raise ValueError("You have to specify either input_ids or inputs_embeds")
+
+        if use_cache and not isinstance(past_key_values, Cache):
+            past_key_values = Cache.from_legacy_cache(past_key_values)
+
+        if inputs_embeds is None:
+            inputs_embeds = self.embeddings(input_ids)
+
+        # embed positions
+        hidden_states = inputs_embeds
+
+        if self.gradient_checkpointing and self.training:
+            if use_cache:
+                logger.warning_once(
+                    "`use_cache=True` is incompatible with gradient checkpointing. Setting `use_cache=False`..."
+                )
+                use_cache = False
+
+        all_hidden_states = () if output_hidden_states else None
+        all_attns = () if output_attentions else None
+        next_cache = None
+
+        for layer in self.layers:
+            if output_hidden_states:
+                all_hidden_states += (hidden_states,)
+
+            if self.gradient_checkpointing and self.training:
+                layer_outputs = self._gradient_checkpointing_func(
+                    layer.__call__,
+                    hidden_states,
+                    attention_mask,
+                    past_key_values,
+                    output_attentions,
+                    use_cache,
+                    **kwargs
+                )
+            else:
+                layer_outputs = layer(
+                    hidden_states,
+                    attention_mask=attention_mask,
+                    past_key_values=past_key_values,
+                    output_attentions=output_attentions,
+                    use_cache=use_cache,
+                    **kwargs
+                )
+
+            hidden_states = layer_outputs[0]
+
+            if use_cache:
+                next_cache = layer_outputs[2 if output_attentions else 1]
+
+            if output_attentions:
+                all_attns += (layer_outputs[1],)
+
+        hidden_states = self.norm(hidden_states)
+
+        # add hidden states from the last decoder layer
+        if output_hidden_states:
+            all_hidden_states += (hidden_states,)
+
+        if not return_dict:
+            return tuple(v for v in [hidden_states, next_cache, all_hidden_states, all_attns] if v is not None)
+
+        return BaseModelOutputWithPast(
+            last_hidden_state=hidden_states,
+            past_key_values=next_cache,
+            hidden_states=all_hidden_states,
+            attentions=all_attns
+        )
+
+
+class TransformerForCausalLM(TransformerPreTrainedModel, GenerationMixin):
+
+    _tied_weights_keys = ["lm_head.weight"]
+
+    def __init__(self, config):
+        super().__init__(config)
+        self.model = TransformerModel(config)
+        self.vocab_size = config.vocab_size
+        self.lm_head = nn.Linear(config.hidden_size, config.vocab_size, bias=False)
+        self.criterion = None
+
+        # Initialize weights and apply final processing
+        self.post_init()
+
+    def get_input_embeddings(self):
+        return self.model.embeddings
+
+    def set_input_embeddings(self, value):
+        self.model.embeddings = value
+
+    def get_output_embeddings(self):
+        return self.lm_head
+
+    def set_output_embeddings(self, new_embeddings):
+        self.lm_head = new_embeddings
+
+    def set_decoder(self, decoder):
+        self.model = decoder
+
+    def get_decoder(self):
+        return self.model
+
+    @deprecate_kwarg("num_logits_to_keep", version="4.50", new_name="logits_to_keep")
+    def prepare_inputs_for_generation(
+        self,
+        input_ids: torch.LongTensor = None,
+        past_key_values: Optional[Union[Cache, List[torch.FloatTensor]]] = None,
+        attention_mask: Optional[torch.Tensor] = None,
+        inputs_embeds: Optional[torch.Tensor] = None,
+        use_cache: bool = True,
+        logits_to_keep: Optional[int] = None,
+        **kwargs
+    ):
+        # only last token for `inputs_ids` if the `past_key_values` is not empty.
+        if past_key_values is not None and len(past_key_values) > 0:
+            input_ids = input_ids[:, -1:]
+        # if `inputs_embeds` are passed, we only want to use them in the 1st generation step
+        if inputs_embeds is not None and len(past_key_values) == 0:
+            model_inputs = {'inputs_embeds': inputs_embeds}
+        else:
+            # The `contiguous()` here is necessary to have a static stride during decoding. torchdynamo otherwise
+            # recompiles graphs as the stride of the inputs is a guard.
+            # Ref: https://github.com/huggingface/transformers/pull/29114
+            # TODO: use `next_tokens` directly instead.
+            model_inputs = {'input_ids': input_ids.contiguous()}
+
+        if logits_to_keep is not None:
+            model_inputs['logits_to_keep'] = logits_to_keep
+
+        model_inputs.update({
+            'past_key_values': past_key_values,
+            'use_cache': use_cache,
+            'attention_mask': attention_mask,
+        })
+        return model_inputs
+
+    @deprecate_kwarg("num_logits_to_keep", version="4.50", new_name="logits_to_keep")
+    def forward(
+        self,
+        input_ids: torch.LongTensor = None,
+        attention_mask: Optional[torch.Tensor] = None,
+        past_key_values: Optional[Union[Cache, List[torch.FloatTensor]]] = None,
+        inputs_embeds: Optional[torch.FloatTensor] = None,
+        labels: Optional[torch.LongTensor] = None,
+        use_cache: Optional[bool] = None,
+        output_attentions: Optional[bool] = None,
+        output_hidden_states: Optional[bool] = None,
+        return_dict: Optional[bool] = None,
+        logits_to_keep: Optional[int] = 0,
+        **kwargs: Unpack[Any]
+    ) -> Union[Tuple, CausalLMOutputWithPast]:
+        output_attentions = output_attentions if output_attentions is not None else self.config.output_attentions
+        output_hidden_states = (
+            output_hidden_states if output_hidden_states is not None else self.config.output_hidden_states
+        )
+        return_dict = return_dict if return_dict is not None else self.config.use_return_dict
+
+        outputs = self.model(
+            input_ids=input_ids,
+            attention_mask=attention_mask,
+            past_key_values=past_key_values,
+            inputs_embeds=inputs_embeds,
+            use_cache=use_cache,
+            output_attentions=output_attentions,
+            output_hidden_states=output_hidden_states,
+            return_dict=return_dict,
+            **kwargs
+        )
+
+        hidden_states = outputs[0]
+        fuse_linear_and_cross_entropy = self.config.fuse_cross_entropy and self.training
+        logits = None if fuse_linear_and_cross_entropy else self.lm_head(hidden_states[:, -logits_to_keep:])
+
+        loss = None
+        if labels is not None:
+            if getattr(self, 'criterion', None) is None:
+                if fuse_linear_and_cross_entropy:
+                    criterion = FusedLinearCrossEntropyLoss()
+                elif self.config.fuse_cross_entropy:
+                    criterion = FusedCrossEntropyLoss(inplace_backward=True)
+                else:
+                    criterion = nn.CrossEntropyLoss()
+            else:
+                criterion = self.criterion
+            # Enable model parallelism
+            labels = labels.to(hidden_states.device)
+            labels = torch.cat((labels[..., 1:], torch.full_like(labels[:, :1], criterion.ignore_index)), 1)
+            labels = labels[..., :hidden_states.shape[1]].contiguous()
+            if fuse_linear_and_cross_entropy:
+                loss = criterion(hidden_states, labels, self.lm_head.weight, self.lm_head.bias)
+            else:
+                loss = criterion(logits.view(labels.numel(), -1), labels.view(-1))
+
+        if not return_dict:
+            output = (logits,) + outputs[1:]
+            return (loss,) + output if loss is not None else output
+
+        return CausalLMOutputWithPast(
+            loss=loss,
+            logits=logits,
+            past_key_values=outputs.past_key_values,
+            hidden_states=outputs.hidden_states,
+            attentions=outputs.attentions,
+        )

fla/models/transformer_mtp/configuration_transformer.py

@@ -0,0 +1,76 @@
+# -*- coding: utf-8 -*-
+
+from typing import Optional
+
+from transformers.configuration_utils import PretrainedConfig
+
+
+class MTPTransformerConfig(PretrainedConfig):
+
+    model_type = 'mtp_transformer'
+    keys_to_ignore_at_inference = ['past_key_values']
+
+    def __init__(
+        self,
+        hidden_size: int = 2048,
+        num_hidden_layers: int = 24,
+        num_heads: int = 32,
+        num_kv_heads: int = None,
+        qkv_bias: bool = False,
+        qk_norm: bool = False,
+        window_size: Optional[int] = None,
+        rope_theta: Optional[float] = 10000.,
+        max_position_embeddings: int = 2048,
+        hidden_ratio: Optional[int] = 4,
+        intermediate_size: Optional[int] = None,
+        hidden_act: str = "swish",
+        initializer_range: float = 0.006,
+        elementwise_affine: Optional[bool] = True,
+        norm_eps: float = 1e-6,
+        use_cache: bool = True,
+        pad_token_id: int = None,
+        bos_token_id: int = 1,
+        eos_token_id: int = 2,
+        tie_word_embeddings: bool = False,
+        fuse_norm: bool = True,
+        fuse_swiglu: bool = True,
+        fuse_cross_entropy: bool = True,
+        vocab_size: int = 32000,
+        n_future_tokens: int = 1,
+        use_custom_backward: Optional[bool] = False,
+        **kwargs,
+    ):
+        self.hidden_size = hidden_size
+        self.num_hidden_layers = num_hidden_layers
+        self.num_heads = num_heads
+        self.num_kv_heads = num_kv_heads
+        self.qkv_bias = qkv_bias
+        self.qk_norm = qk_norm
+        self.window_size = window_size
+        self.rope_theta = rope_theta
+        self.max_position_embeddings = max_position_embeddings
+
+        self.hidden_ratio = hidden_ratio
+        self.intermediate_size = intermediate_size
+        self.hidden_act = hidden_act
+
+        self.initializer_range = initializer_range
+        self.elementwise_affine = elementwise_affine
+        self.norm_eps = norm_eps
+        self.use_cache = use_cache
+
+        self.fuse_norm = fuse_norm
+        self.fuse_swiglu = fuse_swiglu
+        self.fuse_cross_entropy = fuse_cross_entropy
+        self.vocab_size = vocab_size
+
+        self.n_future_tokens = n_future_tokens
+        self.use_custom_backward = use_custom_backward
+
+        super().__init__(
+            pad_token_id=pad_token_id,
+            bos_token_id=bos_token_id,
+            eos_token_id=eos_token_id,
+            tie_word_embeddings=tie_word_embeddings,
+            **kwargs,
+        )